2-bit Open on Hugging Face
Parameters
4.5B
Weights (est.)
≈ 2.0 GB
Mac memory (est.)
8 GB+
Size is estimated from the parameter count and quantisation. Memory allows a fifth extra for context plus 2 GB for the system; long contexts need more.
Run it
With mlx-vlm on an Apple silicon Mac:
pip install mlx-vlmpython -m mlx_vlm.generate --model mlx-community/Ternary-Bonsai-2-4B-2bit --image photo.jpg --prompt "Describe this image"
Details
- Base model
- Qwen/Qwen3.5-4B (finetune) · Qwen3.5 4B on the radar →
- Vendor
- Qwen
- Task
- Vision + text
- Quantisation
- 2-bit
- Licence
- apache-2.0
- Downloads
- 547 in the last 30 days · 547 all time · 0 likes
- Published
- 2026-10-15 · updated 2026-10-15 (26d ago)
- Tags
- prism_hadamard_qwen35, ternary, metal, on-device, hybrid-attention, vision, prismml, bonsai
Other MLX builds of Qwen/Qwen3.5-4B
| Build | Quantisation | Weights | Mac memory | Downloads |
|---|---|---|---|---|
| Qwen3.5-4B-MTP-4bit | 4-bit | ≈ 68 MB | 8 GB+ | 594 |
| Qwen3.5-4B-MTP-5bit | 5-bit | ≈ 83 MB | 8 GB+ | 446 |
| Qwen3.5-4B-MTP-bf16 | bf16 | ≈ 241 MB | 8 GB+ | 201 |
| Ternary-Bonsai-2-4B-2bit (this one) | 2-bit | ≈ 2.0 GB | 8 GB+ | 547 |
| Qwen3.5-4B-OptiQ-4bit | 4-bit | ≈ 2.4 GB | 8 GB+ | 5.5K |
| Qwen3.5-4B-MLX-4bit | 4-bit | ≈ 3.0 GB | 8 GB+ | 102K |
| Qwen3.5-4B-MLX-8bit | 8-bit | ≈ 5.1 GB | 8 GB+ | 1.7K |
| Qwen3.5-4B-MLX-bf16 | bf16 | ≈ 9.1 GB | 16 GB+ | 357 |