Mistral Medium 3.5 (128B)
Mistral AI's frontier dense 128B model with configurable reasoning (instant or chain-of-thought via `reasoning_effort`). Native function-calling, 256K context. Workstation-tier on-device — needs ~80GB of usable memory for Q4_K_M; an MLX 4-bit build is available for Apple Silicon Mac Studio-class hardware. License is a Modified MIT with a revenue carve-out — free for commercial and non-commercial use unless the licensee company exceeds Mistral's revenue threshold.
At a glance
- Parameters
- 128B
- Quantization
- Q4_K_M
- Context window
- 262K tokens
- Approx. size
- 74.9 GB
- Engines
- llama.cpp, MLX
- License
- MIT-Modified (custom-restricted)
- Version
- v1.0.0 , released 2026-05-22
- Family
- mistral
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
0.7- sampling.topP
0.95- sampling.topK
64- sampling.maxTokens
8192- sampling.repetitionPenalty
1.05- sampling.repetitionContext
20- reasoning.thinkingBudget
4096
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- fabrication.detect-past-tense-no-tools
- turn.preamble-folding
- provider.flatten-tool-transcript
Sources
- llama.cpp (GGUF)
- unsloth/Mistral-Medium-3.5-128B-GGUF · Q4_K_M
- MLX
- mlx-community/Mistral-Medium-3.5-128B-4bit
- Upstream
- https://huggingface.co/mistralai/Mistral-Medium-3.5-128B
- Manifest
- View on GitHub
Tags
- mistral
- multimodal
- vision
- reasoning
- tools
- large
- revenue-limited