Mistral Medium 3.5 128B — MLX 4-bit (LM Studio)

Text · Vision · Thinking · Tool Calling

Converted from mistralai/Mistral-Medium-3.5-128B (FP8, 133.6 GB) to MLX 4-bit for Apple Silicon.

Key features

• Vision — Pixtral encoder at full precision (BF16, 48 layers)
• Thinking — Set reasoning_effort to "high" for deep reasoning with [THINK]...[/THINK] blocks
• Tool calling — Native [TOOL_CALLS] tokens
• 256K context — YaRN scaling, RoPE theta 1M

Requirements

• RAM: 72 GB minimum
• Architecture: mistral3
• Quantization: 4-bit affine, group_size=64

Sampling

From Mistral AI and Unsloth recommendations.

Reasoning mode (reasoning_effort="high")

For complex prompts, coding, research, math, and agentic usage:

Parameter	Value
temperature	0.7
top_p	0.95
top_k	20

Non-reasoning mode (reasoning_effort="none")

For fast replies, chat, extraction, and simple instructions:

Parameter	Value
temperature	0.0–0.7
top_p	0.8

Keep repetition and presence penalties disabled (1.0) unless you see looping.

Credits

Role	Author
Original model	Mistral AI
MLX conversion	froggeric

License

Modified MIT — revenue cap $20M/month. See full license.