Ttimms/KAT-Coder-V2.5-Dev-REAP-50-NVFP4A16
Text Generation • 19B • Updated • 2.08k
KAT-Coder-V2.5-Dev · REAP-50 for 16 GB
KAT-Coder-V2.5-Dev pruned 50% (REAP) + quantized for a 16 GB GPU. SWE-bench Verified 52.0%. github.com/t-timms/kat-coder-nvfp4
Note Primary release. 12.45 GiB, NVFP4A16, vLLM. HumanEval+ 90.9%, MBPP+ 89.9%.
Ttimms/KAT-Coder-V2.5-Dev-REAP-50-NVFP4-W4A4
Text Generation • 19B • Updated • 939Note Native FP4 kernel variant — comparison only (~0.84x A16 throughput on SM120).
Ttimms/KAT-Coder-V2.5-Dev-REAP-50-NVFP4A16-GPTQ
Text Generation • 19B • Updated • 92Note GPTQ quant. Documented negative result: no accuracy gain over RTN. Kept for repro.
-
Ttimms/KAT-Coder-V2.5-Dev-REAP-50-GGUF
Text Generation • 19B • Updated • 1.63k • 2 -
Ttimms/KAT-Coder-V2.5-Dev-REAP-50-bf16
Text Generation • 19B • Updated • 1.09k