Qwen3.8-27B-MLX-bf16
MLX BF16 (full precision) version of Qwen/Qwen3.8-27B.
- 4-bit version: lesa80/Qwen3.8-27B-MLX-4bit
All capabilities preserved:
- ✅ Vision — image + video (27-layer ViT encoder)
- ✅ Reasoning — thinking mode (токены
thinking/response) - ✅ Длинный контекст — 262K токенов
- ✅ Multilingual — русский, английский, китайский
| Property | Value |
|---|---|
| Tensor type | bfloat16 |
| Model size | ~51 GB |
| RAM required | 64+ GB |
| License | Apache 2.0 |
Usage
hf download lesa80/Qwen3.8-27B-MLX-bf16 --local-dir ./qwen-mlx-bf16 --local-dir-use-symlinks False
from mlx_vlm import load, generate
model, processor = load("lesa80/Qwen3.8-27B-MLX-bf16")
output = generate(model, processor, "Describe this image.", image=["image.jpg"])
print(output)
mlx_vlm.generate --model lesa80/Qwen3.8-27B-MLX-bf16 --prompt "Hello" --max-tokens 200