Mistral Medium 3.5: the Open Model for Enterprise Agents

1. What Mistral Actually Shipped

Mistral Medium 3.5 is the company’s first flagship merged model: one dense 128B checkpoint with a 256K context window that handles chat, reasoning, and agentic coding without switching models. Mistral trained the vision encoder from scratch for variable image sizes, and reasoning effort is now configurable per request — the same model can answer a quick reply or grind through a long multi-tool run.

Why it matters for enterprise teams: you get an EU vendor with EU data centers, open weights you can self-host on as few as four GPUs, and a hosted API behind the mistral-medium-latest ID. That trio — sovereignty story, exit option, and managed API — is exactly what procurement checklists ask for.

2. Spec Sheet

Everything below comes from Mistral’s official docs card and pricing page, not third-party rumor. This is the ficha you can paste into an architecture decision record.

3. What Improves Over the Previous Generation

Medium 3.5 does not replace the whole lineup — it takes the top agentic slot. Here is how it maps against what came before.

4. Quickstart: Call It in 5 Minutes

You need a La Plateforme API key from console.mistral.ai. The model ID is mistral-medium-latest. Below is the exact pattern from Mistral’s docs — install, chat completion, and the curl equivalent.

pip install mistralai

import os
from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

res = client.chat.complete(
    model="mistral-medium-latest",
    messages=[
        {"role": "system", "content": "You are a senior backend engineer."},
        {"role": "user", "content": "Refactor this function for clarity and add types: def f(d): return [x for x in d if x]"},
    ],
)
print(res.choices[0].message.content)

curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d ’{"model": "mistral-medium-latest", "messages": [{"role": "user", "content": "Hello from the terminal"}]}’

5. Where It Earns Its Price in Production

Medium 3.5 is not a general-chat bargain — it is an agent workhorse. These are the workloads where teams report the premium paying off.

6. Honest Limits

I tested the docs and the pricing page so you don’t have to learn these the expensive way. None of these are dealbreakers — all of them affect your budget or architecture.

7. Verdict

Medium 3.5 is the first Mistral release I’d default to for agent loops instead of chat: 77.6% SWE-Bench Verified in a 128B dense model you can actually self-host, with an EU-hosted API behind it. The $7.5 output price means you deploy it surgically, not everywhere.

Conclusion

Mistral Medium 3.5 consolidates Mistral’s agent story into one 128B open-weight model: 256K context, 77.6% SWE-Bench Verified, $1.5 input / $7.5 output, self-hostable from four GPUs. It replaces Devstral 2 in Vibe and becomes the Le Chat default for a reason — long-horizon reliability.

Start hosted with mistral-medium-latest tonight, measure against Large 3 and Small 4 on your own prompts, and only then decide what to self-host. That’s the whole enterprise playbook in one paragraph.