MLX-Qwopus3.5-27B-v3-vision-5bit

A 5-bit MLX quantization of Jackrong/Qwopus3.5-27B-v3 with a few tweaks to restore the multimodal capabilities.

Supports {%- set enable_thinking = false %} Jinja variable.

Update [2026-04-12]: Refined the chat template to further improve stability for long-running tasks and tool use; mitigated an issue where incorrect think tag formatting could leak from a distillation dataset.

---

Quantization Details

Property	Value
Method	5-bit (5.578 bits per weight)
Tool	mlx-vlm 0.4.2 via mlx-vlm.convert
Size	~19.5GB

---

Other Available Quants

Model	Size	Quantization	Bits per weight	Multimodal
Jackrong/MLX-Qwopus3.5-27B-v3-4bit	15.15 GB	4-bit	4.501	✗
matt-here/MLX-Qwopus3.5-27B-v3-vision-4bit	16.08 GB	4-bit	4.695	✓ (Vision)
matt-here/MLX-Qwopus3.5-27B-v3-5bit	18.56 GB	5-bit	5.501	✗
(This model)	19.46 GB	5-bit	5.678	✓ (Vision)
Jackrong/MLX-Qwopus3.5-27B-v3-6bit	21.88 GB	6-bit	6.501	✗
matt-here/MLX-Qwopus3.5-27B-v3-vision-6bit	22.85 GB	6-bit	6.661	✓ (Vision)
Jackrong/MLX-Qwopus3.5-27B-v3-bf16	53.81 GB	bf16	16	✗

GGUF quants - Jackrong/Qwopus3.5-27B-v3-GGUF

---

Credits

• Alibaba Qwen Team — Qwen 3.5 27B dense model
• Jackrong - Claude 4.6 Opus v3 distillation work
• Unsloth - Training framework
• Apple MLX Team - High-speed local inference on Apple Silicon