Half the thinking, same answers.
If you only have a minute, what you care about is how often the answer is correct and finished within that minute. On one DGX Spark, Qwen3.8-Flash-Next at its default effort spends most of its time thinking, and much of that thinking is second-guessing. Our Qwen3.8-Flash-Next-AZ-Focus does the same work in about half the tokens, with the same answers.
The setup
- Hardware: one NVIDIA DGX Spark (GB10, 128 GB) per model.
- Software, identical for both: Mia's single-Spark recipe on TensorFold's Zig engine, from the recipe's published image, with MTP drafts.
- Base model: the recipe's checkpoint, azampatti's INT4-AutoRound. It routes each token to 5 experts instead of Qwen's 10, for speed, and is healed to stay close to the original. Every comparison below is against that checkpoint.
- Ours: Qwen3.8-Flash-Next-AZ-Focus INT4, the same files except the attention and shared-expert output weights. Point the recipe's
MODEL_IDat it. - Settings: Qwen's default effort (xhigh), the recipe's sampling, and a 128k output limit so no answer was cut short.
1. Same answers, about half the thinking
| Benchmark | Questions | Base | AZ-Focus | Thinking, base | Thinking, AZ-Focus | Change |
|---|---|---|---|---|---|---|
| MMLU-Pro | 500 | 85.4% | 86.4% | 5,880 | 2,849 | -52% |
| BIG-Bench Hard | 540 | 87.6% | 87.4% | 1,962 | 1,059 | -46% |
| GPQA Diamond | 198 | 88.9% | 90.4% | 20,172 | 10,721 | -47% |
| AIME 2025 | 30 × 2 | 98.3% | 98.3% | 27,342 | 17,858 | -35% |
| All 1,298 answers | 1,135 | 1,142 | −46% |
Mean thinking tokens per answer. Single runs of a few hundred questions move by about a point either way. None of these questions was used to build the model.
2. More correct answers in the time you have
The same results, scored the way you experience them: an answer counts only if it is correct and finished within a budget. Tokens are time here: both models generate at about the same speed on the Spark (next section).
| Correct within | 2k tokens | 4k tokens | 8k tokens | 16k tokens | 32k tokens | no limit tokens |
|---|---|---|---|---|---|---|
| Base, all questions | 64% | 68% | 72% | 78% | 82% | 87% |
| AZ-Focus, all questions | 70% | 75% | 79% | 83% | 86% | 88% |
| Base, GPQA + AIME | 34% | 41% | 57% | 70% | 91% | |
| AZ-Focus, GPQA + AIME | 47% | 59% | 71% | 85% | 92% |
On the hardest questions the difference is largest: within 8k tokens AZ-Focus has 59% of them correct and finished, the base model 41%. Given unlimited time both end up in the same place.
3. Time to finish on one Spark
The same 60 questions as our GLM-5.3-Flash post (30 MMLU-Pro, 30 BIG-Bench Hard), at the default effort, one Spark per model, wall clock:
AZ-Focus finished the set 3.3× sooner with one user and 3.7× sooner with four, with the same answers correct. Typical questions take about as long on both; the difference is the long tail, where the base model keeps re-checking an answer it already has. The speed per token is about the same (AZ-Focus about 4% lower: its shorter reasoning is a little harder for the MTP drafter to predict), so the gain is all in writing less.
We also ran RigMark, Alex Ellis's serving benchmark for DGX Spark recipes, on the public INT4 release. Raw speed is the same for both models (51, 81 and about 125 tokens a second for one, two and four users). At the default effort, AZ-Focus finished 2 of RigMark's 3 coding tasks within its 32k output limit, where the base model finished none, and took 22% less decode time over the run. Results and receipts are on the model page.
4. Agentic coding unchanged
We guard agentic coding inside our search, not after it. On SWE-bench Verified (first 100 tasks, mini-SWE-agent, two runs each) AZ-Focus resolved 97 and 96, the base model 97 and 97. Along the way we rejected a version that cut thinking further but started submitting fixes before checking them properly: it scored 92.
5. Code and markup
Asked for SVG, XML, JSON, HTML and CSV (20 prompts, two runs each, every output checked with a strict parser), AZ-Focus wrote about 28% fewer tokens with the same number of valid outputs (35 of 40 for both). Both models sometimes answer an SVG request with a whole HTML page, and Qwen3.8 occasionally writes XML comments containing --, which strict SVG parsers reject: worth knowing whichever version you run.
What this does and doesn't show
- Use the default (xhigh) effort. At
loweffort Qwen3.8 already thinks briefly; AZ-Focus saves almost nothing there and scores slightly lower (102 vs 107 of 128 on our probe). - Set an output limit. On 2 of 1,298 answers AZ-Focus kept exploring past 128k tokens without answering, where the base model answered; on other seeds it solved both, faster than the base.
- Agent work: quality is unchanged, and so is token use: the saving is in reasoning, not in agent steps.
- The original 10-expert model responds differently to the same change, so it has its own release, as EXL3 and BF16: 39% less thinking on our 128-answer set (109 vs 112 correct) and SWE-bench Verified 98 vs 96.
Reproduce it
Everything is public: the models, the recipe, and our receipts (each answer's tokens and correctness for both models, the time-to-complete runs, ttc.py, the 60 questions and the exact commands):
- Benchmarks, 1,298 answers each: base, AZ-Focus
- Time to complete: base, 1 user, AZ-Focus, 1 user, base, 4 users, AZ-Focus, 4 users
- RigMark: base, xhigh, AZ-Focus, xhigh, base, low, AZ-Focus, low
Thanks to Qwen for Qwen3.8-Flash-Next, Intel for the AutoRound quantisation, Saren-Arterius for the hybrid checkpoint and n-gram table, azampatti for the single-Spark checkpoint, Mia for the recipe, Ash Hart for TensorFold, and Alex Ellis for RigMark.