Mistral Releases Medium 3.5: 128B Dense Model With 256k Context and Configurable Reasoning

Mistral AI released Mistral Medium 3.5, a 128B parameter dense model with a 256k context window that handles instruction-following, reasoning, and coding in unified weights.

Model Specifications

• Parameters: 128B (dense architecture)
• Context window: 256k tokens
• Modality: Multimodal input (text and images), text output
• License: Modified MIT (open-source with revenue restrictions)

The model replaces Mistral Medium 3.1 and Magistral in Le Chat, and replaces Devstral 2 in Mistral's coding agent Vibe.

Key Technical Features

Mistral Medium 3.5 introduces configurable reasoning effort, allowing the same model to switch between fast responses and complex reasoning tasks. According to Mistral AI, users can set reasoning_effort="none" for quick replies or reasoning_effort="high" for complex agentic tasks.

The company trained the vision encoder from scratch to handle variable image sizes and aspect ratios. The model supports native function calling and JSON output for agentic applications.

Benchmark Performance

According to Mistral AI:

• τ³-Telecom: 91.4%
• SWE-Bench Verified: 77.6%

Mistral claims the model supersedes all previous Mistral coding models including Devstral across all benchmarks. The company states it achieves "strong results" on instruction following, reasoning (math), and coding benchmarks, though specific scores were not disclosed for all tests.

Language and Capabilities

The model supports dozens of languages including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic. Mistral emphasizes "best-in-class agentic capabilities" and strong adherence to system prompts.

Deployment and Availability

Mistral Medium 3.5 is available through:

• Mistral AI API
• vLLM (recommended for production)
• SGLang
• llama.cpp (text only via Unsloth's GGUF)
• Ollama
• Transformers library

For faster local inference, Mistral released an accompanying EAGLE model for use with vLLM or SGLang. Fine-tuning is supported via Axolotl, Unsloth, and vLLM.

Recommended settings: temperature 0.7 for reasoning_effort="high" and 0.0-0.7 for reasoning_effort="none" depending on task.

What This Means

Mistral's unified architecture approach differs from competitors who maintain separate specialized models. The configurable reasoning represents an implementation of test-time compute scaling, allowing users to trade latency for performance on-demand. At 128B parameters, this sits between mid-tier and frontier models, targeting users who need strong performance without the cost of 400B+ parameter models. The Modified MIT license with revenue restrictions makes this effectively open-weight rather than fully open-source.