Meet Swift
Open-source models with efficient reasoning
6,104,539 downloads on Hugging Face See the breakdown
The models
4 releasesSwift 1.5 Qwen3.8-Flash-Next
63.4% fewer thinking tokens and a 1.8x speed-up, with under 1% accuracy loss.
1,572,238 downloads
Swift 1.5 Qwen3.8-27B
58.5% fewer thinking tokens, scoring 0.35% higher than the base model.
3,318,764 downloads
Swift Bonsai 2 27B
Merged ternary weights in 1-bit and 2-bit GGUF. An experimental research release.
41,098 downloads
Swift-Qwen3.8-27B
58.3% fewer thinking tokens with under 1% accuracy loss.
1,172,439 downloads
Every build
Swift 1.5 Qwen3.8-Flash-Next
- Base modelHugging Face
- AWQ INT4Hugging Face
- AutoRound INT4Hugging Face
- NVFP4Hugging Face
- GGUFHugging Face
- GSQ-RCO GGUFHugging Face
Swift 1.5 Qwen3.8-27B
- Base modelHugging Face
- AWQ INT4Hugging Face
- AutoRound INT4Hugging Face
- AWQ + GPTQ INT4Hugging Face
- NVFP4Hugging Face
- AMD FP8Hugging Face
- GGUFHugging Face
- GSQ-RCO GGUFHugging Face
- MLX 3-bit (text only)Hugging Face
- MLX 4-bitHugging Face
- MLX 5-bitHugging Face
Swift Bonsai 2 27B
- Base modelHugging Face
- 2-bit PQ2_0Hugging Face
- 1-bit PTQ1_0Hugging Face
Swift-Qwen3.8-27B
- Base modelHugging Face
- GGUFHugging Face
- NVFP4Hugging Face
- AMD INT4Hugging Face
- AMD BF16Hugging Face