mistralai/Mistral-Medium-3.5-128B
Mistral Medium 3.5 is a dense 128 billion parameter multimodal language model developed by Mistral AI, featuring a 256k token context window. It excels in instruction-following, complex reasoning, and coding tasks, integrating vision capabilities and configurable reasoning effort per request. This model is designed for unified performance across chat, agentic coding, and general reasoning applications.
Loading preview...
Mistral Medium 3.5: A Unified Multimodal Powerhouse
Mistral Medium 3.5 is a dense 128 billion parameter model from Mistral AI, designed to replace previous models like Mistral Medium 3.1 and Devstral 2. It features an expansive 256k context window and is notable for its multimodal input capabilities, accepting both text and image inputs to produce text outputs. A key innovation is its configurable reasoning effort, allowing the same model to switch between fast chat replies and complex agentic problem-solving, boosting performance with test-time compute when requested.
Key Capabilities
- Unified Performance: Excels across instruction-following, complex reasoning, and coding tasks.
- Multimodal Vision: Analyzes images and provides insights based on visual content, with a newly trained vision encoder for variable image sizes and aspect ratios.
- Agentic Functionality: Offers best-in-class agentic capabilities with native function calling and JSON output, replacing Devstral 2 in the Vibe coding agent.
- Multilingual Support: Supports dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic.
- Strong System Prompt Adherence: Provides robust support for system prompts.
Performance Highlights
Mistral Medium 3.5 demonstrates strong benchmark results, scoring 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified, outperforming previous Mistral coding models. It is now powering Mistral AI's Le Chat application due to its unified capabilities across instruction following, reasoning, and coding benchmarks.
Recommended Usage
Users can configure reasoning_effort to 'none' for quick responses or 'high' for complex prompts and agentic usage, with corresponding temperature and top_p settings. The model is available for inference via vLLM (recommended), SGLang, and Transformers, and supports fine-tuning with Axolotl and Unsloth. It is released under a Modified MIT License.