Mistral AI just released Mistral Medium 3.5, a dense 128-billion-parameter model with a 256k context window that merges instruction-following, reasoning, and coding into a single set of open-weight parameters. It's the company's most capable self-hostable model to date — and it runs on as few as four GPUs.
What Makes Mistral Medium 3.5 Different
Unlike Mistral's previous approach of separate models for different tasks, Medium 3.5 is a "merged" flagship. One model handles chat, complex reasoning, and production coding workflows. The reasoning effort is configurable per request — the same endpoint can fire back a quick reply or work through a multi-step agentic task.
The model replaces both Mistral Medium 3.1 and Magistral in Le Chat, and takes over from Devstral 2 as the engine behind Mistral's coding agent, Vibe.
Key Specs
- Parameters: 128B dense
- Context window: 256,000 tokens
- License: Modified MIT (open weights)
- Modalities: Text and image input, text output
- Self-hosting: As few as 4 GPUs
- Speculative decoding: EAGLE model available for vLLM and SGLang acceleration
Benchmark Performance
Mistral Medium 3.5 posts competitive numbers against much larger models:
- SWE-Bench Verified: 77.6% — ahead of Devstral 2 and Qwen3.5-397B-A17B
- τ³-Telecom: 91.4 — strong agentic capability score
- Multilingual: Dozens of languages including English, French, Spanish, German, Chinese, Japanese, Korean, and Arabic
The vision encoder was trained from scratch to handle variable image sizes and aspect ratios — not bolted on as an afterthought.
Practical Self-Hosting
This is where Medium 3.5 stands out from closed competitors. A dense 128B model on 4 GPUs is achievable for teams that need data sovereignty or want to avoid per-token API costs at scale. The EAGLE speculative decoding model further reduces inference latency on vLLM and SGLang deployments.
For comparison, frontier models like GPT-5.5 and Claude Opus 4.7 offer higher raw capability but require API access. Mistral Medium 3.5 occupies a compelling middle ground — competitive performance with full control over your infrastructure.
Vibe Remote Agents
The release coincides with Mistral's new Vibe remote coding agents. Coding sessions run in the cloud asynchronously, can be spawned from the CLI or Le Chat, and continue working while you step away. Local sessions can be "teleported" to the cloud mid-task, carrying history and state across. Medium 3.5 is the engine that makes long-horizon autonomous coding practical.
Who Should Pay Attention
- Teams self-hosting AI: This is Mistral's best self-hostable model yet. If you're running your own inference stack, it's worth benchmarking against your current setup.
- Developers building agents: The configurable reasoning effort and strong agentic scores make it well-suited for tool-calling workflows.
- Cost-sensitive operations: 128B dense with modified MIT licensing means predictable costs at scale — no per-token surprises.
Getting Started
Mistral Medium 3.5 is available now on Hugging Face and Ollama. It's also the default model in Mistral's Le Chat and Vibe CLI.