Swift 1.5 27B
#14 v Reasoning modelyUkisai · v1.5 · 27b · 3× · naposledy 26. 9. 2026
Swift 1.5 27B (full name: Swift-1.5-Qwen3.8-27b) is a Reasoning model published by ukisai (UkisAI), based on Qwen3.8-27B and designed as a "reasoning-efficient" derivative. It significantly reduces the number of Thinking Tokens compared to the base model while achieving slightly improved or nearly equivalent accuracy, resulting in a noticeable speed advantage. The model is distributed as BF16 weights and in numerous quantization formats (GGUF, AWQ, GPTQ, NVFP4, MLX, among others) via Hugging Face, and is additionally available through a free, API-compatible research interface from ukisai.com. Licensing is based on the Swift Open License v1.0 (free up to $1M annual revenue, thereafter Enterprise License), while the underlying
Vlastnosti
| Key Benchmark (%) | 58.5% fewer thinking tokens while scoring 0.35% higher than base (9.18x speed-up on several tasks) |
| Context Window (Tokens) | 262,144 tokens |
| License | Swift Open License v1.0 (free up to US$1M annual revenue, enterprise license required above); base Qwen3.8-27B under Apache License 2.0 |
| Multimodality | Yes – Image-Text-to-Text (vision+text); text-only variants (e.g., some MLX quants) also available without vision |
| Platform | Hugging Face (BF16 & quants: GGUF, AWQ, GPTQ, NVFP4, MLX); also OpenAI-compatible API at ukisai.com/api/swift/v1 |
| Release Date | September 2026 (Swift 1.5, direct upgrade from Swift 1.0) |