Mistral Medium 3.5

3.65
Mistral has always felt a little different from the big US AI labs. Part of that is the open-model strategy, part of it is the whole “European alternative” angle. But Medium 3.5 is where the company’s priorities become much easier to read.
Advertisement 728 × 90
CompanyMistral AI
Context262K
Released2026-04
Updated2026-09-07

Mistral Medium 3.5 Overview

Mistral released Medium 3.5 on April 29, 2026. It has 128 billion parameters, uses a dense architecture, and supports up to 256K tokens of context.

It is also Mistral’s first “merged model,” meaning instruction following, reasoning, and coding all live in the same set of weights. There is no need to bounce between different models or checkpoints depending on what you are doing.

It handles text, coding and debugging, images, function calls, external tools, and multi-step agent tasks. Language support includes Chinese, Japanese, French, and dozens of others.

The weights are available on Hugging Face under a modified MIT license and can be used for both commercial and non-commercial work.

The benchmark numbers people will probably quote are 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom.

That puts Medium 3.5 in a very different conversation from the Mistral models people once treated as “good for an open model.”

A 77.6% SWE-Bench score is serious.

Still, I would not read too much into it on its own.

SWE-Bench tells you whether a model can solve real GitHub issues under benchmark conditions. It does not tell you how annoying the model will be inside a repository with stale dependencies, bad abstractions, half-documented code, and three failed fixes already sitting in the context.

That gap matters more to me than a few extra leaderboard points.

Mistral Medium 3.5 Pricing

PlanPriceDescription
API input $1.50 / 1M tokens Charged by input tokens
API output $7.50 / 1M tokens Charged by generated tokens

The API supports up to 262,144 context tokens, and the maximum output can also reach 262,144 tokens.

Claude Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens, so Medium 3.5 is roughly half the price.

DeepSeek-V4 Pro is closer on input pricing at around $1.74 per million tokens.

If you are just chatting with a model, these differences may not feel dramatic.

With agents, they do.

A coding agent might read half a repo, call tools, edit files, run tests, inspect the failure, and start again. A few loops like that and token usage climbs fast.

At that point, “half the price” stops sounding like a marketing bullet and starts looking like an engineering constraint.

Mistral Medium 3.5 Key Features

1. Adjustable Reasoning: High Is Useful, Just Not All the Time

Medium 3.5 supports a reasoning_effort setting.

Use "none" and it responds faster. Use "high" and the model spends more compute on the problem.

The difference becomes obvious with code.

For autocomplete, explaining a function, or fixing a small and obvious bug, I would leave reasoning low. High mode often makes you wait longer without giving you much back.

Cross-file changes are a different story.

If the model has to trace dependencies or figure out why some old code was written the way it was, High starts earning its keep. It checks more relationships before committing to a fix.

I noticed this while working through a cross-file issue. The model had already found the source of the error on the first pass, but High reasoning kept going and followed the dependency chain one level deeper.

The final fix was better.

The wait was also very noticeable.

That was the moment the setting clicked for me: High does not always feel smarter. Sometimes it just feels less willing to hand in an answer early.

And that is not necessarily a bad thing.

I just would not pay for that behavior on every request.

2. Native Vision: Fine on Its Own, Much Better Next to Code

Medium 3.5 can work directly with images, including UI screenshots, flowcharts, and images containing equations.

That part did not excite me much by itself. Vision is table stakes for this class of model now.

Then you put it next to the coding features.

A frontend developer can give it a screenshot and the source code in the same task. A debugging workflow can include an error screenshot, an architecture diagram, and the relevant files without handing the job off to another model.

That is where it becomes useful.

The vision system is not the headline feature. It is one more source of context the agent can use while working.

3. Function Calling and Agents: This Is Where the Model Actually Clicks

Function calling and structured JSON output are built in.

With 256K of context, Medium 3.5 can keep code, tool outputs, previous actions, and task history around for fairly long workflows.

Mistral is already using it in its Vibe Coding Agent, where a session can modify code, install dependencies, run tests, and submit pull requests through the CLI or Le Chat.

This is where the whole model starts to make sense.

The long context is there because agents need to keep a lot of state.

reasoning_effort is there because some steps deserve more compute than others.

Function Calling gives the model something to actually do.

Vision lets it see screenshots, interfaces, and diagrams that would otherwise sit outside the loop.

Individually, none of that is surprising.

Together, it feels like Mistral designed Medium 3.5 backward from the question: what does a coding agent actually need to keep working without falling apart after three steps?

That is a much more interesting product decision than simply adding another benchmark point.

Summary

Strengths

  • 128B dense model with coding, vision, reasoning, and tool use in one place
  • 256K context window
  • Maximum output of roughly 262K tokens
  • 77.6% on SWE-Bench Verified
  • Reasoning effort can be adjusted per task
  • API pricing is roughly half of Claude Sonnet 4.6
  • Open weights and self-hosting support
  • The feature set makes much more sense for agents than for plain chat

Limitations

  • No free API tier
  • A 128B dense model is not cheap or easy to self-host
  • High reasoning can spend more time thinking without always producing a better answer
  • 256K context means you can feed it a huge amount of code; it does not mean the model will understand every dependency correctly
  • For everyday chat, summaries, or light writing, most of this is overkill

Who Should Use It?

Software engineers, coding-agent builders, larger AI teams, and companies that care about self-hosting or tighter control over their data.

If you are already building long tool-driven coding workflows, I would test Medium 3.5 on a real repository before making any decision from benchmark charts.

That is where its strengths are supposed to show up.

Who Should Skip It?

People on a tight budget, anyone looking for a free general-purpose chatbot, and users who mainly care about having the highest benchmark score available.

For me, the 77.6% SWE-Bench number is not the reason to care about Medium 3.5.

The more interesting question is whether it can get through a messy coding task with fewer bad edits, fewer wasted tool calls, and less babysitting.

If the answer is yes, the model has a real reason to exist.

If all you need is chat, it is probably too much model.

Comments (0)

Leave a comment

Advertisement 728 × 90

Similar Models

Claude Opus 4.8
100
Claude Opus 4.8 is Anthropic's most powerful Opus-series model, featuring multimodal input, reasoning, and a 1M-token context window, excelling in complex reasoning and coding.
Anthropic
GPT-5.5 Pro
100
I’ve been using GPT-5.5 Pro with a few colleagues for the past several weeks, mostly on the kinds of jobs where regular chatbots tend to fall apart: long documents, messy research questions, code debugging, and tasks that need more than one or two steps of reasoning.
OpenAI
Claude Opus 5
100
What interests me most about Opus 5 isn’t the benchmark bump. It’s that Anthropic has pushed near-flagship capability into a price range more teams can actually use every day. I still wouldn’t treat it as a model you can leave alone for hours and assume everything will work out.
Anthropic
GPT-5.5
99
My first impression of GPT-5.5 is that it feels less like a chatbot and more like a coworker you can actually hand work to. Give it a task and it can break it down, use tools, and check what it has done.
OpenAI
GPT-5.6 Luna Pro
99
OpenAI’s next-generation efficient AI model, balancing speed, cost, and intelligence for chat, coding, and automation tasks.
OpenAI
DeepSeek V4.1 Flash
99
DeepSeek V4.1 Flash has taken over from Pro with a larger model, a 1-million-token context window, built-in vision, and lower API pricing. Some published scores already beat V4 Pro. What matters now is how that holds up in real workloads.
DeepSeek
Gemini 3.8 Flash
98
Gemini 3.8 Flash is a more interesting upgrade than its version number suggests. Google launched Gemini 3.7 Flash just three weeks ago, and now 3.8 Flash is already here. That pace is hard to ignore.
Google
GPT-5.6 Luna
98
An efficient GPT-5.6 series model optimized for speed, cost efficiency, and intelligent performance across large-scale AI applications.
OpenAI