Swift 1.5 Qwen3.8-27B

Demo

We gave the base Qwen3.8-27B and Swift 1.5 27B the same prompt: create a little 3D planet globe a boy scout can walk around while camping, with biomes to explore, small enough to be fun to go around.

The base model took 104.6 minutes to build its game. Swift 1.5 took 11.39 minutes. Play both games here.

Training approach

Starting from Swift 1.0, we scaled up reinforcement learning (RL) and on-policy distillation (OPD) to recover and improve performance, with a focus on long-horizon agent work and coding tasks.

Swift 1.5 keeps the Swift 1.0 recipe, which penalizes the reasoning markers that trigger overthinking, and uses the standard Qwen3.8 interface with text, image, and video support.

Benchmark results

The results below compare Qwen3.8-27B, the foundation base model, Swift 1.0, the first UkisAI adaptation, and Swift 1.5. All three use the same saved evaluation protocols, and every score is a final five-repeat aggregate. Mean-token columns report reasoning tokens per trial; Terminal-Bench sums reasoning across agent calls.

* Note: Terminal Bench 2.1 score of Swift 1.5 is misleading at first glance. It falls in the -38.7% token usage range. Swift has fewer overthinking errors leading to early task failures, leading to longer pursuit of task and higher scores, albeit increasing token usage averages.

Swift 1.5 scores above Swift 1.0 on six of the seven benchmarks reported for both, and above the base model on GPQA-Diamond, C-Eval, LiveCodeBench, and Terminal-Bench 2.1, where it solves 72.13% of tasks against the base model's 69.21%.

Swift 1.5 uses 25–54% fewer thinking tokens than the base model on every question benchmark, with the largest savings on ERQA and C-Eval and the smallest on LiveCodeBench. On Terminal-Bench 2.1 it uses 16.3% fewer reasoning tokens on average, with the median almost unchanged.

Reasoning effort

Qwen3.8's reasoning_effort setting lets users choose how much the model thinks. For Swift 1.5 to be useful across these settings, it needs to reduce thinking while keeping accuracy close to the base, so we tested xhigh, medium, and low. Thinking-token savings persist at every level; at low, Swift 1.5 scores above the base while using about 29% fewer thinking tokens.

Quantized models

Available Swift 1.5 27B formats are grouped below. The benchmark table after this list covers only the three INT4 exports evaluated under the stated protocol; it does not report results for the other formats.

These results evaluate the merged Swift 1.5 checkpoint and three INT4 exports on GPQA-Diamond (198 questions), IFBench (300 prompts), and AIME 2026 (30 problems). Each model completed the full datasets with one sample per prompt, seed 0, and zero request errors. This is a single-seed evaluation, separate from the five-repeat BF16 results above.

The Qwen base columns use saved seed-0 runs. Quantization recipes and serving settings differ from the new Swift 1.5 runs, so these are reference comparisons rather than a controlled measurement of the Swift adaptation; the token reductions are recomputed from those same reference samples. For AIME, truncated responses count as incorrect in both columns. The AMD Quark INT4 and FP8 exports have separate sanity evaluations; completed results on these three benchmarks are not available for them.

Run Swift 1.5

Swift 1.5 is served through an OpenAI-compatible API at ukisai.com/api/swift/v1, free for research and without an API key; the model id is swift. To run it yourself, serve the full model with vLLM or SGLang, using a build that supports Qwen3.8, and adjust tensor parallelism and context length to your GPU memory.

vllm serve ukisai/Swift-1.5-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 8000

python -m sglang.launch_server \
  --model-path ukisai/Swift-1.5-Qwen3.8-27b \
  --dtype bfloat16 \
  --tp-size 1 \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --port 8000

curl https://ukisai.com/api/swift/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "swift", "messages": [{"role": "user", "content": "Hello, Swift."}]}'

# vLLM
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'

# SGLang
--speculative-algorithm EAGLE --speculative-num-steps 3 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

For hardware-specific options, see the vLLM recipe or SGLang recipe. The published weights include the base model's MTP head for self-speculative decoding; the model card has the full instructions.

License

Swift 1.5 is a derivative of Qwen3.8-27B (Apache License 2.0). UkisAI's contribution, including the adapted weights, is licensed 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. Above that threshold, commercial use requires a separate Swift Enterprise License. Nothing in the Swift Open License limits rights in Qwen3.8-27B itself under Apache 2.0.

Contact us for enterprise licensing.