Swift-Qwen3.8-27B
#33 in Open-Source LLMsUkisai · v3.8 · 27b · 2× · last seen Sep 18, 2026
Swift-Qwen3.8-27B is a reasoning-efficient Fine-Tuning derivative of Alibaba/Qwen's open Qwen3.8-27B model, released by UkisAI. It reduces the number of Thinking-Tokens generated during Reasoning by up to 58.3% while maintaining nearly unchanged accuracy (averaging less than 1% loss), achieving up to 1.95x speed improvements on certain tasks. The model inherits text, image, and video input capabilities from the base model and is distributed under the proprietary "Swift Open License v1.0" with a revenue threshold for commercial use. According to independent analysis, gains are not consistent across all Benchmarks: certain tests like AIME 2026 or HMMT show slight accuracy losses, while others such as LiveCodeBench v6 show improvements.
Features
| Key Benchmark (%) | GPQA-Diamond: -0.1 percentage points vs. base model; 58.3% fewer thinking tokens with <1% average performance loss |
| Context Window (Tokens) | 262,144 tokens |
| License | Swift Open License v1.0 (free for personal/research/commercial use up to US$1,000,000 annual revenue; Swift Enterprise License required above that); base Qwen3.8-27B remains Apache 2.0 |
| Multimodality | Text, image, and video input (inherited from the Qwen3.8-27B base model) |
| Platform | Hugging Face (Safetensors, GGUF, NVFP4, INT4-AMD variants); usable via vLLM, SGLang, llama.cpp, Ollama; own OpenAI-compatible API at ukisai.com/api/swift/v1 |
| Price per 1M Tokens | Free research API (OpenAI-compatible, limited to 5 requests/minute, no API key required); no commercial per-token price published by UkisAI |
| Release Date | Initial model release approx. 6 days ago, announcement blog post approx. 1 week ago (as of research date) |