Training approach
Reasoning models often repeat checks or revisit answers they have already found. Swift reduces this extra reasoning.
We identified tokens associated with overthinking in Qwen’s reasoning and penalized their use during training. Swift produces shorter reasoning traces, and we observe fewer overthinking errors in our testing.
Swift includes a transfer component derived from BottleCap AI’s ThinkingCap-Qwen3.6-27B. The resulting model uses the standard Qwen3.8 interface and retains text, image, and video support.
Benchmark results
We evaluated Qwen3.8-27B and Swift on nine benchmarks, with BF16 weights and five runs per model. All base scores come from our own evaluations.
| Benchmark | Accuracy | Mean tokens | Token reduction | |||
|---|---|---|---|---|---|---|
| Base | Swift | Base | Swift | Mean | Median | |
| General reasoning | ||||||
| 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% |
| Mathematics | ||||||
| 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% |
| Multimodal | ||||||
| ERQA | 67.45% | 66.30% | 4,137 | 2,045 | ↓ 50.6% | ↓ 54.6% |
| Coding & agents | ||||||
| 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% |
Swift uses fewer tokens on every benchmark. Accuracy stays close on GPQA and C-Eval and improves on LiveCodeBench. AIME and HMMT show larger accuracy losses, as the table shows.
How to reproduce
Serving configuration
BF16 · vLLM 0.27.1 · Qwen3 reasoning parser · context 262,144. Thinking is enabled at xhigh. Sampling: temperature=1.0, top_p=0.95, top_k=20, min_p=0, presence_penalty=0, repetition_penalty=1.
Benchmark configurations
Five request seeds (0–4) per model, with averages reported across runs. Terminal-Bench uses five trials per task; IFBench uses strict scoring. LiveCodeBench covers the full release_v6.
| Benchmark | Output cap |
|---|---|
| GPQA-Diamond | 100,000 |
| MMLU-Pro | 100,000 |
| C-Eval | 16,384 |
| IFBench | 81,920 |
| AIME 2026 | 250,000 |
| HMMT Nov 2025 | 250,000 |
| ERQA | 100,000 |
| Terminal-Bench 2.1 | Agent/task limits |
| LiveCodeBench v6 | 32,768 |
Mean reduction is computed from the base and Swift token means; median reduction compares their median token counts. The headline 58.3% refers to the GPQA-Diamond median reduction, the largest in the suite. LiveCodeBench measures completion tokens; the other rows measure thinking tokens. Terminal-Bench counts tokens per complete trial.
Some base outputs were reused from saved runs. Exact replay requires the original dataset snapshots and harness manifests, which are retained internally.
Reasoning effort
Qwen3.8 lets users set reasoning_effort. We tested medium and low as well as xhigh; Swift uses fewer thinking tokens at all three settings.
| Reasoning effort | Mean thinking reduction |
|---|---|
| Xhigh | ↓ 41.0% |
| Medium | ↓ 22.7% |
| Low | ↓ 25.8% |
Effort comparison settings
These are matched BF16 measurements from one benchmark, repeated across five seeds at each effort level, with the same 100,000-token output cap. They establish token savings at each setting; they do not establish unchanged accuracy across the full suite at medium and low effort.
Quantized models
Swift is intended for quantized deployment. We tested mixed-precision W4A16 and AWQ INT4 weights and found that token savings carry over.
| Benchmark / quantization | Base accuracy | Swift accuracy | Mean reduction | Median reduction |
|---|---|---|---|---|
| GPQA-DiamondMixed-precision W4A16 | 88.69% | 88.38% | ↓ 32.1% | ↓ 50.2% |
| IFBenchMixed-precision W4A16 | 72.58% | 71.25% | ↓ 30.1% | ↓ 38.0% |
| AIME 2026Mixed-precision W4A16 | 84.00% | 84.00% | ↓ 19.0% | ↓ 37.5% |
| AIME 2026AWQ INT4 | 82.67% | 84.00% | ↓ 22.8% | ↓ 34.8% |
On AIME, Swift matches or improves accuracy and reduces output-cap failures by 31–33%. GPQA accuracy stays close to the quantized base; IFBench accuracy is slightly lower.
Quantized evaluation settings
GPQA and AIME use five seeds; IFBench uses four samples per prompt and strict scoring. Output caps: GPQA 100,000; IFBench 81,920; AIME 32,768. GPQA measures thinking tokens; IFBench and AIME measure completion tokens. GPQA and IFBench use historical base runs. AIME uses template-default effort and counts truncated answers as incorrect, so it is a separate comparison from the BF16 math results.
Run Swift
The full model is available on Hugging Face. Serve it with vLLM or SGLang, using a build that supports Qwen3.8. Adjust tensor parallelism and context length to your GPU memory.
vLLM configuration
vllm serve ukisai/Swift-Qwen3.8-27b \
--dtype bfloat16 \
--tensor-parallel-size 1 \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--port 8000SGLang configuration
python -m sglang.launch_server \
--model-path ukisai/Swift-Qwen3.8-27b \
--dtype bfloat16 \
--tp-size 1 \
--context-length 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder \
--port 8000For hardware-specific options, see the vLLM recipe or SGLang recipe. The model’s MTP head is included; optional decoding flags are in the model card.
License
Swift is released under the Swift Open License v1.0. Personal, research, educational, evaluation, and commercial use are free for individuals and organizations with gross annual revenue, including affiliates, of up to US$1 million. Commercial use above that threshold requires a Swift Enterprise License.
Acknowledgements
We acknowledge the NVIDIA Innovation Lab for providing access to 8× NVIDIA H100 GPUs to train Swift.