M

Mistral Medium 3.5

Free
4.4
71
New AI ToolsFreeFree tier
Inputs: text, imageOutputs: text
Type
Saas
Company
Mistral AI

About Mistral Medium 3.5

Mistral Medium 3.5 is a 128-billion-parameter dense language model developed by Mistral AI, publicly available under a modified MIT open-weights license. Designed as a flagship merged model, it handles instruction-following, reasoning, and coding tasks through a single set of weights, with a 256,000-token context window. The model also includes a vision encoder trained from scratch to process variable image sizes and aspect ratios. It is optimized for agentic workflows, long-horizon tasks, and structured output generation, achieving notable scores such as 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom. Mistral Medium 3.5 is deployable on as few as four GPUs for self-hosted use, and its reasoning effort can be configured per request, allowing it to switch between quick responses and complex multi-step operations.

This model serves as the default engine for two major Mistral AI products: Le Chat and Mistral Vibe. In Le Chat, Mistral Medium 3.5 powers a new 'Work mode' that enables multi-step tasks such as research, analysis, and cross-tool actions by calling tools in parallel until completion. In Mistral Vibe, it supports remote asynchronous coding agents that run in the cloud, can be spawned from the command-line interface or directly from Le Chat, and continue working while users step away. These agents are designed for async coding sessions that offload tasks without interrupting conversations.

As part of Mistral AI's broader ecosystem, Mistral Medium 3.5 represents a convergence of coding, reasoning, and instruction-following capabilities into a single compact model. Its open-weights release allows developers to self-host and fine-tune the model, while hosted access is available through Mistral's platform. The platform itself includes additional features such as text and image input processing, though current capabilities are focused on text and image modalities. Pricing for hosted services and free-tier limits are not explicitly detailed in the available scraped content and should be verified on the official website.

Key Features

128-billion-parameter dense model with 256,000-token context window
Open weights released under a modified MIT license
Configurable reasoning effort per request
Vision encoder handling variable image sizes and aspect ratios
Optimized for agentic tasks and structured output generation
Deployable on as few as four GPUs for self-hosting
Powers remote async coding agents in Mistral Vibe and Work mode in Le Chat

Pros & Cons

Pros
  • Open-weights availability enables self-hosting and customization
  • Strong benchmark performance (77.6% on SWE-Bench Verified, 91.4 on τ³-Telecom)
  • Large 256K context window for handling long documents or conversations
  • Configurable reasoning effort allows flexible use from quick replies to deep analysis
  • Runs efficiently on relatively few GPUs (four) for a 128B model
  • Integrated into Mistral's Le Chat and Vibe platforms for seamless user experience
Cons
  • Model is in public preview; stability and feature completeness should be verified
  • Hosted service free tier and usage limits are not clearly stated and should be checked
  • Self-hosting a 128B model still requires substantial hardware and expertise
  • Output quality may vary depending on prompt engineering and task complexity
  • Vision capabilities are present but the full multimodal feature set might be limited compared to dedicated vision models

Best For

Coding assistance and automated code generationComplex multi-step reasoning and problem-solvingAsynchronous coding tasks that run in the cloud without user supervisionResearch and analysis involving cross-tool actionsBuilding agentic workflows that require tool calling and structured outputSelf-hosted AI inference for enterprise or privacy-sensitive applications

Alternatives to Mistral Medium 3.5

FAQ

Is Mistral Medium 3.5 free to use?
The model weights are released under a modified MIT license, which permits free use, modification, and distribution. However, hosted access via Mistral's platform may have its own pricing; free-tier availability and limits should be verified on the official website.
What hardware is needed to run Mistral Medium 3.5 locally?
The model can be self-hosted on as few as four GPUs. Exact GPU requirements (e.g., VRAM and type) are not provided in the given content and should be checked in the official documentation.
Does Mistral Medium 3.5 support image inputs?
Yes, the model includes a vision encoder trained from scratch to handle variable image sizes and aspect ratios, enabling image understanding.
How does Work mode in Le Chat differ from the standard chat?
Work mode is powered by Mistral Medium 3.5 and designed for complex, multi-step tasks. It calls tools in parallel until the task is complete, whereas standard chat is more suited for one-off queries.
Is Mistral Medium 3.5 available for commercial use?
The model is released under a modified MIT license, which generally permits commercial use, but users should review the exact terms of the license for any restrictions.
Can I use Mistral Medium 3.5 in my own coding agent?
Yes, the model is designed for agentic tasks and produces structured output. It can be integrated into custom agents, and Mistral Vibe provides a ready-to-use remote coding agent infrastructure.