Generate Perplexity, KLD, ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-F16.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 42 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-F16.gguf (version GGUF V3 (latest))
+
+Final result: 61.7333 +/- 1.7759
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =   40261.05 ms
+llama_perf_context_print: prompt eval time =  149462.93 ms / 36666 tokens (    4.08 ms per token,   245.32 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  150294.55 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-F16.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 42 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-F16.gguf (version GGUF V3 (latest))
+
+750	79.73333333%	[76.7082%, 82.4554%]
+
+
+llama_perf_context_print:        load time =    2970.89 ms
+llama_perf_context_print: prompt eval time =  528861.23 ms / 129319 tokens (    4.09 ms per token,   244.52 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  532325.38 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-F16.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 42 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-F16.gguf (version GGUF V3 (latest))
+
+Final result: 40.8000 +/- 1.7958
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =    2768.82 ms
+llama_perf_context_print: prompt eval time =  276273.33 ms / 68956 tokens (    4.01 ms per token,   249.59 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  277471.13 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-F16.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 42 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-F16.gguf (version GGUF V3 (latest))
+
+Final result: 37.4667 +/- 1.7686
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =    2730.94 ms
+llama_perf_context_print: prompt eval time =  213745.70 ms / 51053 tokens (    4.19 ms per token,   238.85 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  215254.19 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-F16.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 42 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-F16.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 72.4000 +/- 1.6334
+
+llama_perf_context_print:        load time =    2841.75 ms
+llama_perf_context_print: prompt eval time =   91265.69 ms / 22541 tokens (    4.05 ms per token,   246.98 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   91746.92 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_M.gguf (version GGUF V3 (latest))
+
+Final result: 65.3333 +/- 1.7389
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    4590.43 ms
+llama_perf_context_print: prompt eval time =  152258.69 ms / 36666 tokens (    4.15 ms per token,   240.81 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  153118.42 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_M.gguf (version GGUF V3 (latest))
+
+750	80.00000000%	[76.9876%, 82.7066%]
+
+
+llama_perf_context_print:        load time =     769.34 ms
+llama_perf_context_print: prompt eval time =  538472.19 ms / 129319 tokens (    4.16 ms per token,   240.16 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  542167.13 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_M.gguf (version GGUF V3 (latest))
+
+Final result: 39.6000 +/- 1.7870
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     806.33 ms
+llama_perf_context_print: prompt eval time =  283405.69 ms / 68956 tokens (    4.11 ms per token,   243.31 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  284720.03 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  25.645736 ±   0.214748
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  95.06%
+Mean ln(PPL(Q)/PPL(base))     :   0.226040 ±   0.002601
+Mean PPL(Q)/PPL(base)         :   1.253625 ±   0.003260
+Mean PPL(Q)-PPL(base)         :   5.188479 ±   0.079257
+
+====== KL divergence statistics ======
+Mean    KLD:   0.286497 ±   0.002130
+Maximum KLD:  27.942949
+99.9%   KLD:  11.911683
+99.0%   KLD:   3.204565
+99.0%   KLD:   3.204565
+Median  KLD:   0.139660
+10.0%   KLD:   0.006744
+ 5.0%   KLD:   0.001423
+ 1.0%   KLD:   0.000043
+Minimum KLD:  -0.000894
+
+====== Token probability statistics ======
+Mean    Δp: -0.836 ± 0.032 %
+Maximum Δp: 99.942%
+99.9%   Δp: 76.283%
+99.0%   Δp: 35.932%
+95.0%   Δp: 16.691%
+90.0%   Δp:  9.040%
+75.0%   Δp:  1.129%
+Median  Δp: -0.019%
+25.0%   Δp: -2.256%
+10.0%   Δp: -11.765%
+ 5.0%   Δp: -20.807%
+ 1.0%   Δp: -43.493%
+ 0.1%   Δp: -79.148%
+Minimum Δp: -99.887%
+RMS Δp    : 12.354 ± 0.058 %
+Same top p: 77.725 ± 0.107 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_M.gguf (version GGUF V3 (latest))
+
+Final result: 35.0667 +/- 1.7436
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     759.94 ms
+llama_perf_context_print: prompt eval time =  216548.55 ms / 51053 tokens (    4.24 ms per token,   235.76 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  218211.64 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_m.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_M.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 70.9333 +/- 1.6591
+
+llama_perf_context_print:        load time =     753.88 ms
+llama_perf_context_print: prompt eval time =   94337.41 ms / 22541 tokens (    4.19 ms per token,   238.94 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   94843.38 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_S.gguf (version GGUF V3 (latest))
+
+Final result: 62.2667 +/- 1.7711
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    4272.29 ms
+llama_perf_context_print: prompt eval time =  152740.59 ms / 36666 tokens (    4.17 ms per token,   240.05 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  153595.41 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_S.gguf (version GGUF V3 (latest))
+
+750	78.00000000%	[74.8968%, 80.8179%]
+
+
+llama_perf_context_print:        load time =     728.94 ms
+llama_perf_context_print: prompt eval time =  537974.27 ms / 129319 tokens (    4.16 ms per token,   240.38 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  541689.42 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_S.gguf (version GGUF V3 (latest))
+
+Final result: 39.4667 +/- 1.7860
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     757.43 ms
+llama_perf_context_print: prompt eval time =  283333.69 ms / 68956 tokens (    4.11 ms per token,   243.37 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  284662.93 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  32.512225 ±   0.267969
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  91.81%
+Mean ln(PPL(Q)/PPL(base))     :   0.463278 ±   0.003282
+Mean PPL(Q)/PPL(base)         :   1.589276 ±   0.005215
+Mean PPL(Q)-PPL(base)         :  12.054968 ±   0.135871
+
+====== KL divergence statistics ======
+Mean    KLD:   0.484943 ±   0.002308
+Maximum KLD:  23.137739
+99.9%   KLD:  11.445929
+99.0%   KLD:   3.996119
+99.0%   KLD:   3.996119
+Median  KLD:   0.279736
+10.0%   KLD:   0.019254
+ 5.0%   KLD:   0.004941
+ 1.0%   KLD:   0.000332
+Minimum KLD:  -0.000675
+
+====== Token probability statistics ======
+Mean    Δp: -4.988 ± 0.042 %
+Maximum Δp: 99.950%
+99.9%   Δp: 75.785%
+99.0%   Δp: 36.459%
+95.0%   Δp: 13.799%
+90.0%   Δp:  5.484%
+75.0%   Δp:  0.086%
+Median  Δp: -0.427%
+25.0%   Δp: -7.430%
+10.0%   Δp: -24.702%
+ 5.0%   Δp: -37.934%
+ 1.0%   Δp: -63.532%
+ 0.1%   Δp: -87.413%
+Minimum Δp: -99.733%
+RMS Δp    : 17.076 ± 0.062 %
+Same top p: 70.771 ± 0.117 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_S.gguf (version GGUF V3 (latest))
+
+Final result: 36.1333 +/- 1.7553
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     755.36 ms
+llama_perf_context_print: prompt eval time =  216405.85 ms / 51053 tokens (    4.24 ms per token,   235.91 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  217987.98 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq3_s.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ3_S.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 72.9333 +/- 1.6235
+
+llama_perf_context_print:        load time =     752.81 ms
+llama_perf_context_print: prompt eval time =   94369.26 ms / 22541 tokens (    4.19 ms per token,   238.86 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   94874.43 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ4_NL.gguf (version GGUF V3 (latest))
+
+Final result: 62.4000 +/- 1.7699
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    5546.40 ms
+llama_perf_context_print: prompt eval time =  157820.59 ms / 36666 tokens (    4.30 ms per token,   232.33 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  158658.47 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ4_NL.gguf (version GGUF V3 (latest))
+
+750	79.46666667%	[76.4290%, 82.2040%]
+
+
+llama_perf_context_print:        load time =     856.94 ms
+llama_perf_context_print: prompt eval time =  558886.79 ms / 129319 tokens (    4.32 ms per token,   231.39 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  562503.08 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ4_NL.gguf (version GGUF V3 (latest))
+
+Final result: 41.4667 +/- 1.8002
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     914.88 ms
+llama_perf_context_print: prompt eval time =  294610.76 ms / 68956 tokens (    4.27 ms per token,   234.06 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  295846.99 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  25.450410 ±   0.199570
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  96.78%
+Mean ln(PPL(Q)/PPL(base))     :   0.218394 ±   0.001994
+Mean PPL(Q)/PPL(base)         :   1.244077 ±   0.002481
+Mean PPL(Q)-PPL(base)         :   4.993153 ±   0.059469
+
+====== KL divergence statistics ======
+Mean    KLD:   0.173444 ±   0.001426
+Maximum KLD:  22.223246
+99.9%   KLD:   9.213535
+99.0%   KLD:   1.556305
+99.0%   KLD:   1.556305
+Median  KLD:   0.081647
+10.0%   KLD:   0.006194
+ 5.0%   KLD:   0.001743
+ 1.0%   KLD:   0.000131
+Minimum KLD:  -0.000788
+
+====== Token probability statistics ======
+Mean    Δp: -3.757 ± 0.027 %
+Maximum Δp: 99.782%
+99.9%   Δp: 44.427%
+99.0%   Δp: 18.257%
+95.0%   Δp:  5.991%
+90.0%   Δp:  2.080%
+75.0%   Δp:  0.027%
+Median  Δp: -0.328%
+25.0%   Δp: -5.093%
+10.0%   Δp: -15.863%
+ 5.0%   Δp: -24.394%
+ 1.0%   Δp: -43.233%
+ 0.1%   Δp: -70.697%
+Minimum Δp: -99.732%
+RMS Δp    : 11.165 ± 0.051 %
+Same top p: 82.789 ± 0.097 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ4_NL.gguf (version GGUF V3 (latest))
+
+Final result: 37.4667 +/- 1.7686
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     912.73 ms
+llama_perf_context_print: prompt eval time =  225621.49 ms / 51053 tokens (    4.42 ms per token,   226.28 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  227248.28 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-iq4_nl.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-IQ4_NL.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 71.3333 +/- 1.6523
+
+llama_perf_context_print:        load time =     937.74 ms
+llama_perf_context_print: prompt eval time =   98116.26 ms / 22541 tokens (    4.35 ms per token,   229.74 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   98670.87 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_L.gguf (version GGUF V3 (latest))
+
+Final result: 64.0000 +/- 1.7539
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    5136.60 ms
+llama_perf_context_print: prompt eval time =  164152.27 ms / 36666 tokens (    4.48 ms per token,   223.37 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  164972.95 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_L.gguf (version GGUF V3 (latest))
+
+750	77.20000000%	[74.0633%, 80.0595%]
+
+
+llama_perf_context_print:        load time =     833.00 ms
+llama_perf_context_print: prompt eval time =  580747.74 ms / 129319 tokens (    4.49 ms per token,   222.68 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  584554.96 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_L.gguf (version GGUF V3 (latest))
+
+Final result: 41.4667 +/- 1.8002
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     788.56 ms
+llama_perf_context_print: prompt eval time =  305965.89 ms / 68956 tokens (    4.44 ms per token,   225.37 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  307240.49 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  24.405174 ±   0.195423
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  95.33%
+Mean ln(PPL(Q)/PPL(base))     :   0.176457 ±   0.002431
+Mean PPL(Q)/PPL(base)         :   1.192984 ±   0.002901
+Mean PPL(Q)-PPL(base)         :   3.947917 ±   0.064108
+
+====== KL divergence statistics ======
+Mean    KLD:   0.248285 ±   0.001640
+Maximum KLD:  22.227032
+99.9%   KLD:   9.707264
+99.0%   KLD:   2.304967
+99.0%   KLD:   2.304967
+Median  KLD:   0.123584
+10.0%   KLD:   0.007649
+ 5.0%   KLD:   0.001796
+ 1.0%   KLD:   0.000084
+Minimum KLD:  -0.000635
+
+====== Token probability statistics ======
+Mean    Δp: -2.090 ± 0.032 %
+Maximum Δp: 99.189%
+99.9%   Δp: 64.212%
+99.0%   Δp: 31.519%
+95.0%   Δp: 13.797%
+90.0%   Δp:  6.831%
+75.0%   Δp:  0.560%
+Median  Δp: -0.057%
+25.0%   Δp: -3.331%
+10.0%   Δp: -14.319%
+ 5.0%   Δp: -24.226%
+ 1.0%   Δp: -48.172%
+ 0.1%   Δp: -79.381%
+Minimum Δp: -99.228%
+RMS Δp    : 12.606 ± 0.056 %
+Same top p: 78.472 ± 0.106 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_L.gguf (version GGUF V3 (latest))
+
+Final result: 36.8000 +/- 1.7621
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     841.25 ms
+llama_perf_context_print: prompt eval time =  233752.65 ms / 51053 tokens (    4.58 ms per token,   218.41 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  235346.12 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_l.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_L.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 72.8000 +/- 1.6260
+
+llama_perf_context_print:        load time =     832.18 ms
+llama_perf_context_print: prompt eval time =  101421.35 ms / 22541 tokens (    4.50 ms per token,   222.25 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  101943.45 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 64.4000 +/- 1.7496
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    4355.26 ms
+llama_perf_context_print: prompt eval time =  158252.70 ms / 36666 tokens (    4.32 ms per token,   231.69 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  159087.45 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_M.gguf (version GGUF V3 (latest))
+
+750	77.60000000%	[74.4798%, 80.4389%]
+
+
+llama_perf_context_print:        load time =     744.01 ms
+llama_perf_context_print: prompt eval time =  561047.56 ms / 129319 tokens (    4.34 ms per token,   230.50 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  564754.58 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 40.8000 +/- 1.7958
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     759.93 ms
+llama_perf_context_print: prompt eval time =  295554.48 ms / 68956 tokens (    4.29 ms per token,   233.31 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  296870.54 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  24.750046 ±   0.198446
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  94.92%
+Mean ln(PPL(Q)/PPL(base))     :   0.190490 ±   0.002538
+Mean PPL(Q)/PPL(base)         :   1.209842 ±   0.003071
+Mean PPL(Q)-PPL(base)         :   4.292789 ±   0.068077
+
+====== KL divergence statistics ======
+Mean    KLD:   0.273478 ±   0.001716
+Maximum KLD:  23.756779
+99.9%   KLD:   9.973213
+99.0%   KLD:   2.569659
+99.0%   KLD:   2.569659
+Median  KLD:   0.139089
+10.0%   KLD:   0.008610
+ 5.0%   KLD:   0.002078
+ 1.0%   KLD:   0.000103
+Minimum KLD:  -0.000674
+
+====== Token probability statistics ======
+Mean    Δp: -2.189 ± 0.033 %
+Maximum Δp: 99.919%
+99.9%   Δp: 63.247%
+99.0%   Δp: 33.168%
+95.0%   Δp: 14.581%
+90.0%   Δp:  7.284%
+75.0%   Δp:  0.603%
+Median  Δp: -0.062%
+25.0%   Δp: -3.479%
+10.0%   Δp: -15.050%
+ 5.0%   Δp: -25.583%
+ 1.0%   Δp: -50.625%
+ 0.1%   Δp: -81.793%
+Minimum Δp: -99.620%
+RMS Δp    : 13.182 ± 0.057 %
+Same top p: 77.344 ± 0.107 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 36.9333 +/- 1.7635
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     738.68 ms
+llama_perf_context_print: prompt eval time =  225672.28 ms / 51053 tokens (    4.42 ms per token,   226.23 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  227200.49 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_m.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_M.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 72.9333 +/- 1.6235
+
+llama_perf_context_print:        load time =     750.12 ms
+llama_perf_context_print: prompt eval time =   97927.96 ms / 22541 tokens (    4.34 ms per token,   230.18 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   98426.78 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_S.gguf (version GGUF V3 (latest))
+
+Final result: 61.0667 +/- 1.7816
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    4042.36 ms
+llama_perf_context_print: prompt eval time =  158459.88 ms / 36666 tokens (    4.32 ms per token,   231.39 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  159304.81 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_S.gguf (version GGUF V3 (latest))
+
+750	77.86666667%	[74.7577%, 80.6916%]
+
+
+llama_perf_context_print:        load time =     704.36 ms
+llama_perf_context_print: prompt eval time =  558138.31 ms / 129319 tokens (    4.32 ms per token,   231.70 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  561898.16 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_S.gguf (version GGUF V3 (latest))
+
+Final result: 40.9333 +/- 1.7967
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     707.14 ms
+llama_perf_context_print: prompt eval time =  294106.51 ms / 68956 tokens (    4.27 ms per token,   234.46 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  295434.68 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  25.982356 ±   0.208609
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  92.33%
+Mean ln(PPL(Q)/PPL(base))     :   0.239080 ±   0.003120
+Mean PPL(Q)/PPL(base)         :   1.270080 ±   0.003963
+Mean PPL(Q)-PPL(base)         :   5.525099 ±   0.086044
+
+====== KL divergence statistics ======
+Mean    KLD:   0.410016 ±   0.002211
+Maximum KLD:  26.456850
+99.9%   KLD:  11.191044
+99.0%   KLD:   3.826428
+99.0%   KLD:   3.826428
+Median  KLD:   0.221591
+10.0%   KLD:   0.014583
+ 5.0%   KLD:   0.003518
+ 1.0%   KLD:   0.000202
+Minimum KLD:  -0.000602
+
+====== Token probability statistics ======
+Mean    Δp: -2.759 ± 0.040 %
+Maximum Δp: 99.843%
+99.9%   Δp: 77.836%
+99.0%   Δp: 42.001%
+95.0%   Δp: 18.458%
+90.0%   Δp:  8.931%
+75.0%   Δp:  0.665%
+Median  Δp: -0.102%
+25.0%   Δp: -4.673%
+10.0%   Δp: -19.427%
+ 5.0%   Δp: -31.918%
+ 1.0%   Δp: -59.329%
+ 0.1%   Δp: -87.266%
+Minimum Δp: -99.664%
+RMS Δp    : 15.977 ± 0.062 %
+Same top p: 72.822 ± 0.114 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_S.gguf (version GGUF V3 (latest))
+
+Final result: 38.2667 +/- 1.7759
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     689.04 ms
+llama_perf_context_print: prompt eval time =  225077.84 ms / 51053 tokens (    4.41 ms per token,   226.82 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  226715.57 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q3_k_s.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q3_K_S.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 72.0000 +/- 1.6406
+
+llama_perf_context_print:        load time =     737.04 ms
+llama_perf_context_print: prompt eval time =   98097.25 ms / 22541 tokens (    4.35 ms per token,   229.78 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =   98614.54 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 62.0000 +/- 1.7736
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    5492.30 ms
+llama_perf_context_print: prompt eval time =  164320.56 ms / 36666 tokens (    4.48 ms per token,   223.14 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  165122.13 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_M.gguf (version GGUF V3 (latest))
+
+750	79.33333333%	[76.2895%, 82.0782%]
+
+
+llama_perf_context_print:        load time =     898.60 ms
+llama_perf_context_print: prompt eval time =  582056.25 ms / 129319 tokens (    4.50 ms per token,   222.18 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  585789.58 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 41.7333 +/- 1.8018
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     923.69 ms
+llama_perf_context_print: prompt eval time =  306593.84 ms / 68956 tokens (    4.45 ms per token,   224.91 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  307849.04 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.ppx

@@ -0,0 +1,37 @@
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  21.411111 ±   0.170166
+Mean PPL(base)                :  20.457257 ±   0.161275
+Cor(ln(PPL(Q)), ln(PPL(base))):  98.21%
+Mean ln(PPL(Q)/PPL(base))     :   0.045572 ±   0.001499
+Mean PPL(Q)/PPL(base)         :   1.046627 ±   0.001569
+Mean PPL(Q)-PPL(base)         :   0.953854 ±   0.032590
+
+====== KL divergence statistics ======
+Mean    KLD:   0.092796 ±   0.001164
+Maximum KLD:  25.177353
+99.9%   KLD:   8.122986
+99.0%   KLD:   0.871791
+99.0%   KLD:   0.871791
+Median  KLD:   0.036392
+10.0%   KLD:   0.001979
+ 5.0%   KLD:   0.000420
+ 1.0%   KLD:   0.000007
+Minimum KLD:  -0.000817
+
+====== Token probability statistics ======
+Mean    Δp: -0.463 ± 0.019 %
+Maximum Δp: 99.023%
+99.9%   Δp: 44.662%
+99.0%   Δp: 19.835%
+95.0%   Δp:  9.041%
+90.0%   Δp:  4.857%
+75.0%   Δp:  0.673%
+Median  Δp: -0.004%
+25.0%   Δp: -1.191%
+10.0%   Δp: -6.370%
+ 5.0%   Δp: -11.362%
+ 1.0%   Δp: -24.949%
+ 0.1%   Δp: -50.867%
+Minimum Δp: -99.932%
+RMS Δp    :  7.244 ± 0.048 %
+Same top p: 87.132 ± 0.086 %

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.tqa

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_M.gguf (version GGUF V3 (latest))
+
+Final result: 39.2000 +/- 1.7838
+Random chance: 19.8992 +/- 1.4588
+
+
+llama_perf_context_print:        load time =     924.45 ms
+llama_perf_context_print: prompt eval time =  234356.12 ms / 51053 tokens (    4.59 ms per token,   217.84 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  235973.86 ms / 51054 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_m.wng

@@ -0,0 +1,19 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_M.gguf (version GGUF V3 (latest))
+
+Final Winogrande score(750 tasks): 71.4667 +/- 1.6500
+
+llama_perf_context_print:        load time =     897.50 ms
+llama_perf_context_print: prompt eval time =  101897.65 ms / 22541 tokens (    4.52 ms per token,   221.21 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  102441.25 ms / 22542 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_s.arc

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_S.gguf (version GGUF V3 (latest))
+
+Final result: 62.9333 +/- 1.7648
+Random chance: 25.0083 +/- 1.5824
+
+
+llama_perf_context_print:        load time =    5073.64 ms
+llama_perf_context_print: prompt eval time =  163620.09 ms / 36666 tokens (    4.46 ms per token,   224.09 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  164444.66 ms / 36667 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_s.hsw

@@ -0,0 +1,20 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_S.gguf (version GGUF V3 (latest))
+
+750	78.40000000%	[75.3141%, 81.1964%]
+
+
+llama_perf_context_print:        load time =     844.26 ms
+llama_perf_context_print: prompt eval time =  578922.97 ms / 129319 tokens (    4.48 ms per token,   223.38 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  582667.59 ms / 129320 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)

scores/Mistral-Small-3.2-24B-Instruct-2506-pruned-q4_k_s.mmlu

@@ -0,0 +1,21 @@
+build: 5890 (982e3472) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
+llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
+llama_model_loader: loaded meta data with 46 key-value pairs and 345 tensors from ./Mistral-Small-3.2-24B-Instruct-pruned-Q4_K_S.gguf (version GGUF V3 (latest))
+
+Final result: 40.4000 +/- 1.7930
+Random chance: 25.0000 +/- 1.5822
+
+
+llama_perf_context_print:        load time =     898.56 ms
+llama_perf_context_print: prompt eval time =  305282.54 ms / 68956 tokens (    4.43 ms per token,   225.88 tokens per second)
+llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_perf_context_print:       total time =  306510.96 ms / 68957 tokens
+ggml_metal_free: deallocating
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
+ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)