HelixML

Swift 1.5 Flash-Next Cut Qwen3.8's Thinking Tokens in Our Pilot

Sep 27, 2026

We deployed UkisAI's Swift 1.5 Qwen3.8-Flash-Next on four RTX PRO 6000 GPUs and compared completed answers, thinking tokens, tool use and serving speed with our previous Qwen checkpoint.

We replaced the Qwen checkpoint on four RTX PRO 6000 Blackwell Server Edition GPUs with UkisAI's Swift 1.5 Flash-Next NVFP4 checkpoint. The OpenAI-compatible model name stayed qwen3.8-flash-next, so existing clients kept working. The other four GPUs continued serving their own models during the change.

UkisAI reports that its BF16 model uses fewer thinking tokens and finishes reasoning tasks faster than base Qwen3.8-Flash-Next. We tested the quantized build we could run in production. A fixed-length tokens-per-second test showed little difference; completed answers told us more.

Two completed math problems

We gave Swift and the previous NVIDIA Qwen NVFP4 checkpoint the same two AIME 2026 questions, with xhigh reasoning, temperature 0, seed 7 and a 32,768-token output cap. Both models answered 2/2 correctly. Across the pair, Swift finished in 476 seconds versus 666 seconds, a 1.40× gain, and used 34,840 versus 48,921 thinking tokens, a 28.8% reduction.

Two matched AIME 2026 questions: both Qwen and Swift answered correctly. Question 10 took 409 seconds on Qwen and 395 on Swift; question 12 took 257 seconds on Qwen and 82 on Swift. Swift's time gain varied from 1.04 to 3.15 times.

Our measurements on the same four GPUs and request settings. Each bar starts at zero; the selected pair is too small for a general speed estimate.

Across the same two completed questions, Qwen used 48,921 thinking tokens and Swift used 34,840, a 28.8 percent reduction. This is a selected two-question pilot, not a population estimate.

Thinking tokens reported by the serving API for those two completed answers. Both models answered both questions correctly.

The gain varied sharply between questions: 1.04× on question 10 and 3.15× on question 12. We selected this pair after pilot runs, and an additional Swift pilot reached a 16,000-token cap without an answer. Two completed questions cannot establish a general speedup or reproduce UkisAI's reported 1.8× result on its larger BF16 evaluation. The baseline's first timed question may also include first-use compilation; we warmed question 12 before measuring it.

Serving checks before rollout

We verified the pinned checkpoint files and serving arguments before loading the GPUs. On the isolated four-GPU engine, 5/5 agent and tool cases passed, as did seven multi-turn and context checks, including synthetic 32K and 128K recall, plus an image request. The fixed 256-token code workload ran at 74.8 versus 74.5 output tokens/s for one request and 474.5 versus 469.3 tokens/s for eight simultaneous requests, Swift versus the prior Qwen deployment. Those tests measure generation rate, not time to finish an open-ended task.

The model checkpoint, UkisAI's model page and our earlier Qwen3.8-Flash-Next serving post provide the model and deployment context.