Quick Answer: Mistral Medium 3.5 is a 128-billion parameter open-weight AI model dropped by Mistral AI in late April 2026. It rolls reasoning, coding, instruction-following and image understanding into a single model — replacing three separate tools the company had been running before.
What Is Mistral Medium 3.5?
If you’ve been following the AI space lately. You already know things move fast. Mistral AI the French company quietly building one of the most competitive AI stacks in Europe. Just made a move that simplifies everything they’ve released so far.
On April 28, 2026, Mistral launched Medium 3.5. It’s not just a new model. It’s a consolidation. Before this, developers had to juggle Mistral Medium 3.1 for general chat, Magistral for reasoning tasks and Devstral 2 for coding. That’s three different models, three different integrations, three different things to manage.
Medium 3.5 kills all that. One model. Every task.
What Can It Actually Do?
Let’s get into the specs — but in plain terms not a press release.
| Feature | What It Means |
|---|---|
| 128B Parameters (Dense) | Every parameter fires on every query — more reliable than models that only activate part of themselves |
| 256,000-Token Context | You can paste an entire codebase or legal document and it won’t lose the thread |
| Text + Image Input | Send a screenshot, a chart, a photo — it handles visuals alongside text |
| Adjustable Reasoning | Need a quick answer? Keep it light. Complex task? Crank up the reasoning effort |
| Runs on 4 GPUs | You don’t need a data center to self-host this |
| Modified MIT License | Free to download and build on commercially, within a revenue threshold |
| API Pricing | $1.50 per million input tokens, $7.50 per million output tokens |
A dense model is worth highlighting here. Unlike mixture-of-experts models that only activate a slice of their parameters on each task. Medium 3.5 uses all 128 billion every single time. That consistency is what makes it predictable in production environments something enterprises genuinely care about.
The Benchmark Numbers
Numbers only mean so much, but these are worth knowing:
- SWE-Bench Verified: 77.6% — this test throws real GitHub issues at the model and checks if it can actually fix them autonomously. 77.6% is a strong result for an open-weight model.
- τ³-Telecom: 91.4% — a technical reasoning benchmark focused on telecom-domain problem solving, where Medium 3.5 beat every previous Mistral model.
For context, these scores place Medium 3.5 in the same ballpark as several closed-source frontier models — at a fraction of the API cost.
Cloud Coding Agents in Vibe CLI
Here’s where things get practically useful for developers.
Mistral’s Vibe CLI tool used to run coding tasks on your local machine. That’s fine for small jobs, but anything large would slow your system down or tie it up for hours. With Medium 3.5, Vibe now offloads work to remote cloud agents that run asynchronously.
The workflow is straightforward:
- Kick off a task from Vibe CLI or directly inside Le Chat
- The cloud agent picks it up and runs it remotely
- Multiple tasks run in parallel — no waiting in line
- You get notified when it’s done
- Review the output and approve anything sensitive before it goes live
That last point matters. The agent won’t send an email, push code, or delete a file without your green light first. It’s autonomous, but not reckless.
Le Chat Work Mode: Your AI Coworker
Le Chat is Mistral’s chat interface — think of it like their version of ChatGPT. The new Work Mode takes it well beyond a chatbot.
Powered by Medium 3.5, Work Mode connects to your real tools — email, Slack, GitHub, Notion and executes multi-step tasks from a single instruction. Tell it to summarize last week’s Slack threads and draft a report? It does that. Ask it to review open GitHub issues and flag urgent ones? Done.
It’s live in preview right now across Free, Pro and Team plans. That the free tier gets access is notable. Most competitors gate agentic features behind paid plans.
Open Weights: Why It Matters
The model weights are publicly available on Hugging Face. You can download them, run them locally, build products on top of them. The Modified MIT License allows commercial use for most businesses, with restrictions kicking in above a certain revenue threshold.
In a market where GPT-4o and Claude Sonnet remain locked behind APIs, having a model at this capability level available as open weights is genuinely significant. It puts serious AI power in the hands of startups, researchers and solo developers who couldn’t otherwise afford frontier-tier performance.
At roughly 70 GB in 4-bit quantization. It’s approaching territory where well-equipped consumer machines can run it — no cloud required.
Who Should Actually Use This?
- Developers building coding tools, autonomous agents, or AI-powered apps
- Enterprises wanting frontier-level performance without frontier-level API bills
- Researchers who need long-context document processing and image understanding
- Startups ready to self-host and avoid vendor lock-in
- Non-technical users who want Le Chat’s Work Mode to handle repetitive business tasks
Frequently Asked Questions
Q: What models does Mistral Medium 3.5 replace?
It replaces Mistral Medium 3.1, the Magistral reasoning model, and Devstral 2 — three products merged into one.
Q: Can I run Mistral Medium 3.5 on my own hardware?
Yes. NVIDIA confirmed it runs on as few as four GPUs. In 4-bit quantization, the model sits around 70 GB — manageable for high-end workstations.
Q: How does it compare to GPT-4o?
On SWE-Bench coding tests, Medium 3.5 scores 77.6% — competitive with GPT-4o. While being cheaper per token and available as open weights.
Q: Is it free to use?
The API is paid ($1.50/M input tokens). The model weights are free to download. Le Chat Work Mode is accessible on the free plan during preview.
Q: When was Mistral Medium 3.5 released?
April 28, 2026.
Expert Take
Mistral Medium 3.5 isn’t just a better model. It’s a smarter product decision. Consolidating three tools into one reduces friction for developers and makes the platform easier to recommend to non-technical stakeholders. The open-weight release at this capability level keeps Mistral relevant in a market increasingly dominated by American closed-source labs. And the cloud agents in Vibe CLI, paired with Work Mode in Le Chat, show that Mistral is building a productivity stack not just shipping models. For teams evaluating AI infrastructure in 2026, Medium 3.5 is worth a serious look.
