AxonZeta/Qwen3.8-Flash-Next-AZ-Focus-INT4

Qwen3.8-Flash-Next-AZ-Focus (INT4, one DGX Spark)

Qwen3.8-Flash-Next with the second-guessing turned down: about half the thinking at Qwen's default effort, with the same accuracy, and agentic coding unchanged.

Qwen3.8-Flash-Next thinks carefully, but on hard questions a lot of that thinking is hesitation: "Wait…", "Hmm, maybe…", re-checking answers it already had. We optimised the model's weights against that habit, while leaving alone the checking it does as an agent (running tests, inspecting output before it submits). AZ-Focus still reasons step by step. It just loops less.

• About half the thinking at the default effort (xhigh): 46% fewer thinking tokens across 1,298 benchmark answers, with accuracy level (1,142 vs 1,135 correct).
• More right answers when time is short: within 8k tokens, 79% vs 72% of answers are right and finished; on the hardest questions (GPQA, AIME) 59% vs 41%.
• Agentic coding unchanged: SWE-bench Verified 97 and 96 / 100 vs 97 and 97 for the base model (first 100 tasks, two runs).
• Code and markup generation about 28% shorter, with output validity the same as the base model.
• Weights only. No adapter, runtime patch or extra code. A drop-in for the same serving setup.

This repository is built on azampatti/Qwen3.8-Flash-Next-125B-A5B-INT4-AutoRound, the single-Spark checkpoint used by Mia's TensorFold recipe. That checkpoint routes each token to 5 experts instead of 10 and is healed to stay close to the original (see its card). All comparisons below are against that checkpoint, on the same recipe and hardware. Only the attention and shared-expert output weights differ from it; the layout, file names and everything else are the same.

Made by AxonZeta, an independent AI lab in Manchester, UK.

Run it on one DGX Spark

With Mia's single-Spark TensorFold recipe, point it at this repository:

MODEL_ID=AxonZeta/Qwen3.8-Flash-Next-AZ-Focus-INT4 MODEL_REVISION=5d2f89941ae201fa3458515a6ab06dd51c25e60b ./start.sh

Everything else (drafts, KV cache, context, effort levels) works as in the recipe. Speed per token is the same as the base checkpoint; answers finish sooner because there are fewer tokens. In a side-by-side on two of our Sparks (the recipe's tools/bench.py, models swapped between machines, thinking on), both ran at 54-56 tok/s for one request, about 190 tok/s for eight at once, and about 2,500 tok/s prefill from 6k to 100k tokens. The recipe's published figures (64.4 tok/s for one request, thinking off) come from a different Spark and settings; compare like with like.

Results (default effort = xhigh)

Benchmark	Base checkpoint	AZ-Focus	Thinking tokens
MMLU-Pro (500)	85.4% · 5,880	86.4% · 2,849	−52%
BIG-Bench Hard (540)	87.6% · 1,962	87.4% · 1,059	−46%
GPQA Diamond (198)	88.9% · 20,172	90.4% · 10,721	−47%
AIME 2025 (30, 2 runs)	98.3% · 27,342	98.3% · 17,858	−35%
All 1,298 answers	1,135 correct	1,142 correct	−46%
SWE-bench Verified, first 100 (mini-SWE-agent, two runs)	97 / 97	97 / 96

Correct within a token budget (an answer counts only if it is right and finished within the budget):

Budget	2k	4k	8k	16k	32k	no limit
Base checkpoint, all questions	64%	68%	72%	78%	82%	87%
AZ-Focus, all questions	70%	75%	79%	83%	86%	88%
Base checkpoint, GPQA + AIME		34%	41%	57%	70%	91%
AZ-Focus, GPQA + AIME		47%	59%	71%	85%	92%

Thinking on, the recipe's sampling (temperature 1.0, top-p 0.95, top-k 20), 128k output limit, same server and hardware for both. Single runs of a few hundred questions move by about a point either way. None of these questions was used to build the model.

Code and markup generation (20 prompts × 2: SVG, XML, JSON, HTML, CSV, each output checked with a strict parser): AZ-Focus used 7,916 tokens per reply vs 11,054 (−28%), with 5 of 40 outputs failing the check for both models. Both sometimes answer an SVG request with a full HTML page, and Qwen3.8 occasionally writes XML comments containing --, which strict SVG parsers reject.

Effort setting and limits

• Use it at the default (xhigh) effort. That is where the hesitation is. At low effort the model already thinks briefly; there AZ-Focus saves almost nothing (3,450 vs 3,596 tokens per answer) and scores slightly below the base (102 vs 107 of 128). If you run at low effort, use the base checkpoint.
• Agentic coding: quality is unchanged, and so is the token use; the saving is on reasoning, not on agent steps.
• Set an output limit. On 2 of 1,298 benchmark answers (one AIME problem, one GPQA puzzle), AZ-Focus kept exploring approaches past 128k tokens without answering, where the base model answered; on other seeds it solved both, in fewer tokens than the base. The base model had no such cases. A max_tokens limit (the recipe's default is 32,768) bounds this.
• Other languages, very long contexts, vision and multi-hour agent sessions haven't been measured separately yet.

License

Qwen Community License 1.0 (see LICENSE), as for the original model. Note its terms for very large products and for Model-as-a-Service / AI work-assistant businesses.

Credits

• Qwen for Qwen3.8-Flash-Next
• Intel for the AutoRound int4 quantization
• Saren-Arterius for the hybrid checkpoint and the fp8 n-gram table
• azampatti for the 5-expert, healed single-Spark checkpoint this is built on
• Mia's AI Lab for the single-Spark recipe, and Ash Hart for TensorFold

Questions, or want to build on this with us? hello@axonzeta.ai