UkisAI's Swift-Qwen3.8-27B sells a median, not a mean
UkisAI's Swift-Qwen3.8-27B claims 58.3% fewer thinking tokens on GPQA-Diamond. That figure is a median, not a mean, and the accuracy loss lands on hard math.
Vlad Makarovreviewed and published
UkisAI published Swift-Qwen3.8-27B on Hugging Face in early September, a post-trained derivative of Qwen3.8-27B whose entire pitch is that it stops thinking sooner. The model card opens with "58.3% fewer thinking tokens", an accuracy loss under one percent and a 1.95x speed-up. The artifact is real: 27.8 billion BF16 parameters in eighteen shards at about 55.6 GB, with the MTP head kept for self-speculative decoding. What the headline numbers are not is audited, and the most-quoted of them is a median rather than a mean.
The headline is a median, and here is the mean
On GPQA-Diamond, UkisAI's own table shows mean thinking tokens falling from 15,014 to 8,855, a 41.0 percent reduction. The 58.3 percent is the median on that same benchmark. Both columns are printed in the card; the abbreviation still travels further than the row it came from. The gap matters: the typical question lost well over half its trace while the long tail shortened far less. Serving capacity is priced on the mean, since memory and wall-clock time scale with the sum rather than the midpoint. A 41 percent cut is respectable. It is not 58.3.
Everything in this table was measured by the seller
Every figure below is self-reported, comparing the same base model with and without the Swift adapter at BF16, vLLM 0.27.1, a 262,144-token context, thinking at xhigh, temperature 1.0 and top_p 0.95, five seeds per model. No third party has reproduced these runs, and no Hugging Face inference provider serves the model.
| Benchmark | Base accuracy | Swift accuracy | Mean tokens (base → Swift) | Mean change | Median change |
|---|---|---|---|---|---|
| GPQA-Diamond | 88.38 | 88.28 | 15,014 → 8,855 | −41.0% | −58.3% |
| MMLU-Pro | 85.47 | 84.95 | 2,980 → 1,603 | −46.2% | −28.3% |
| C-Eval | 90.00 | 90.62 | 1,492 → 804 | −46.1% | −19.3% |
| IFBench | 73.53 | 71.80 | 8,052 → 4,657 | −42.2% | −50.5% |
| AIME 2026 | 98.67 | 94.00 | 22,014 → 16,143 | −26.7% | −50.2% |
| HMMT (Nov 2025) | 99.33 | 96.00 | 22,032 → 15,189 | −31.1% | −45.9% |
| ERQA (multimodal) | 67.45 | 66.30 | 4,137 → 2,045 | −50.6% | −54.6% |
| Terminal-Bench 2.1 | 66.74 | 65.84 | 37,086 → 27,272 | −26.5% | −38.7% |
| LiveCodeBench v6 | 76.76 | 81.55 | 11,374 → 8,615 | −24.3% | −45.8% |
The losses are not spread evenly. Swift gives ground on competition mathematics, the rows where the base sits nearest its ceiling: AIME 2026 drops 4.67 points, HMMT Nov 2025 drops 3.33, and a 4.79-point gain on LiveCodeBench v6 does not offset them. So "under 1% loss" holds only if you weight the rows. Unweighted, the three largest movements are a 4.79-point gain on a coding benchmark, a 4.67-point loss on AIME and a 3.33-point loss on HMMT — and the losses land exactly where a cheaper thinking budget is normally sold as a trade, at the top of the mathematics curve.
The savings survive the effort dial, at a discount
A shorter trace that helps only at maximum effort would be a narrow product, so UkisAI tested all three of the base model's reasoning_effort levels and reports mean savings at each — though the lower settings give back much of the headline.
| Reasoning effort | Mean thinking-token reduction |
|---|---|
| xhigh | 41.0% |
| medium | 22.7% |
| low | 25.8% |
The comparison that matters is against the base model's own dial. On GPQA-Diamond, Swift at xhigh scores 88.28 at 8,855 mean tokens, the base at xhigh scores 88.38 at 15,014, and the base at medium scores 84.14 at 4,451. So Swift buys maximum-effort accuracy at half the tokens without replacing the dial, because plain medium is still cheaper. UkisAI concedes the point: its adapter is "complementary" to the existing effort settings, and reasoning length should be "optimized" rather than "shortened by force".
The method is distillation with a borrowed component
UkisAI identified reasoning-marker tokens its analysis associates with overthinking in Qwen's rollouts and fine-tuned the base to penalize them while it reasons, restoring accuracy with on-policy distillation plus a transfer component from BottleCap AI's ThinkingCap-Qwen3.6-27B. Its announcement frames the target as anxiety-like reasoning loops and cites a paper on overthinking in post-training quantization, which finds quantized models disproportionately sampling markers such as "wait" and "alternatively" at high-entropy positions. That shorter reasoning helps quantized models is established. Whether a token penalty transfers as cleanly as claimed is not.
Running it, and the license that limits it
Quantized deployment is the intended use, and the numbers agree: W4A16 GPQA-Diamond loses 0.31 points while mean tokens fall 32.1 percent, W4A16 AIME 2026 holds at 84.00 with a 19.0 percent cut, and AWQ INT4 AIME improves from 82.67 to 84.00. AIME output-cap failures fall by 31 to 33 percent.
Access splits two ways. A free research endpoint sits at https://ukisai.com/api/swift/v1, model id swift, no API key, five requests per minute, no stated continuity guarantee. Self-hosting is documented: GGUF builds from Q1 to Q8, a bartowski quantization and community NVFP4 and W4A16 variants, plus vLLM or SGLang recipes. The weights carry the Swift Open License v1.0: free for personal, research, educational and commercial use up to one million dollars of annual recurring revenue including affiliates, above which a separate enterprise license applies. That is not OSI-approved open source, and the threshold is a business model. Teams above the line should price the license before the GPUs.
A different story from the other Qwen3.8-27B derivatives
Elsewhere in this family, the community uncensored variants, a humanlike-chat LoRA and a 100K-context local setup change what the model says. Swift changes how long it takes to say it, a claim paid for on an electricity bill, and one that binds on a high-memory used-server build where throughput rather than capacity is the limit.
What would settle it
Two experiments would settle this. The first is a third-party re-run under the published sampling settings, especially on the AIME row, where a 26 percent mean reduction is much weaker than the 58.3 percent in the title. The second is an independent check that the reduction survives at low effort and long context, where a token penalty trained on reasoning traces has the least to work with. The weights are downloadable today; the number doing the marketing, the number that sizes the cluster and the number in the table are three different numbers, and only the vendor has run them.
