Mistral Large 4 Guide: Europe’s 1-Trillion-Parameter Open AI Is Free to Download — Code Fixes, Security Checks, and the $0.70 API Price

In one sentence: imagine Europe’s strongest open AI brain — a model with about a trillion knobs inside it, trained from scratch in European data centers — being handed to the public for free download, while you can also just rent it by the word for less than a dollar per million words in.

Two terms, plainly: open-weight means the company published the model’s actual files, so anyone can download, inspect, and run them — as opposed to a closed model you can only rent through an API. Parameters are the adjustable knobs a model learns during training; more knobs usually means it can hold more patterns, but also means it needs bigger computers to run. Reasoning mode is a dial for how hard the model thinks before answering: Low for quick jobs, High for the hardest problems.

Mistral AI announced the open weights on October 29, 2026, in its official post “Mistral Large 4 (Le Chonk) is now open weight.” The model itself first appeared in public preview on October 6; today’s news is that the weights are out, the model got better after three more weeks of training, and the API price dropped. Its nickname is “Le Chonk.”

Why it matters to you

• Frontier coding help without a frontier bill. Mistral says Large 4 leads open-weight models on coding tests and matches the best open models on agent benchmarks while using far fewer tokens per task (vendor claims — see below). What this means for you: a coding assistant that used to cost flagship prices now runs at $0.70 per million input tokens.
• A dial for cost vs. quality. The new Low / Medium / High reasoning modes let you pay less thinking for easy jobs and reserve full effort for hard ones. What this means for you: you stop paying flagship effort on a typo fix.
• Security work you can actually sell. Mistral positions this model for defensive cybersecurity: reading vulnerability reports, reproducing flaws in a sandbox, and writing patches. What this means for you: small dev teams that cannot afford a security hire are a real buyer for that service.
• No vendor lock-in if you grow. Because the weights are public, a business that starts on the API can later host the model itself. What this means for you: the tool you build today is not held hostage by one company’s price changes tomorrow.

The numbers, with vendor claims labeled

All benchmark figures below are Mistral’s own claims from its October 29 post, not independently verified:

• Size: 1.05 trillion parameters total, about 52 billion active per token — think of it as a huge team where only a small crew works on each word. Native image input (a 1.6-billion-parameter vision encoder) and a 1-million-token context window — roughly, it can hold an entire large codebase in one conversation.
• Coding: 19.7% on FrontierCode 2.0 (a private test of whether a repo maintainer would merge the AI’s change); 41% on Terminal-Bench 4.0, up 13 points from the preview, with 30% fewer output tokens; 48% on the Artificial Analysis Coding Agent Index, effectively tied with the leading open model GLM-5.3 while using 47% fewer tokens per task.
• Defensive security: 51% on the Artificial Analysis Cyber Index (find, reproduce, and patch vulnerabilities without breaking things); 97.1% resistance on the B3 Agent Security Benchmark, up from 93.3% in preview; 96.6% average refusal on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm.
• Work tasks: 73.3% on Finch (finance spreadsheet workflows, second only to Kimi K3 among open models); 84.5% on BrowseComp (web research), up from 75% at its earlier checkpoint.
• Human taste: in Surge Code’s blind side-by-side tests, reviewers preferred Large 4 over other leading open models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated.

Price: $0.70 in, $2.10 out

Per Mistral’s October 29 post (vendor price, not independently verified):

Item	Price per 1M tokens
Input	$0.70
Cached input	$0.07
Output	$2.10

What this means for you: a typical bug-fix session — say 20,000 input tokens of code plus 3,000 output tokens of fixed code — costs roughly $0.02. The model itself costs $0 to download; what you pay for is either the API, or the servers to run it yourself. This replaces the preview API price of $1.36 in / $0.14 cached / $4.18 out — about half price.

One license catch, in plain words: the weights are free under Mistral’s Ecosystem Open-Weight license (a modified MIT license) to use, modify, fine-tune, and commercialize — except that if your business is selling access to the model itself (Model-as-a-Service), you need a separate commercial license from Mistral first. Selling services built with the model (like the bug-fix gig below) is the normal case and does not trigger that clause.

Try it in 5 steps (no download needed)

Use Mistral’s Vibe Code — its coding chat — for the fastest first run. No GPU, no install.

Step 1. Go to chat.mistral.ai (Mistral’s chat app, where Vibe Code lives) and sign in or create a free Mistral account.

Step 2. Start a new coding session and make sure the model is set to Mistral Large 4. If you see a reasoning setting, leave it on High (the default) for your first hard task, or drop it to Low for a quick question — that is the new cost dial.

Step 3. Paste a real piece of code with a real bug, plus this copy-paste prompt:

“Find the bug in this code. Explain in plain English what is wrong and why, then give me the fixed version with the changed lines marked. Do not change anything else.”

Step 4. Read the explanation first, then the fix. Check the changed lines against your own understanding — never ship an AI fix you have not read. If the fix looks wrong, reply: “That breaks X. Try again, keeping the existing behavior for Y.”

Step 5. What you should see: a short plain-English diagnosis, then fixed code with changes marked. That loop — paste, diagnose, verify — is the whole service below.

Prefer the API instead? Sign up at console.mistral.ai, create an API key, and call model mistral-large-4 with a reasoning mode of Low, Medium, or High. The weights themselves are on Hugging Face for teams that run vLLM or SGLang servers, and Microsoft Foundry and NVIDIA NIM also serve the model — those paths are for server teams, not for your first try.

The money play: the bug-fix sprint service

• Who pays: indie hackers, small dev teams, and startups with a backlog of bugs and no spare engineer.
• What you sell: “Send me the bug report and the code. Within 48 hours you get the fix, a plain-English explanation of what was wrong, and a test suggestion — $49 to $149 per fixed bug to start.”
• What to charge: start at $49 per bug for your first 3 clients (proof beats price early), then $99–$149 once you have before/after samples. Your model cost per bug is pennies (see the $0.02 example above), so the margin is your review time.
• 7-day validation: Days 1–2, fix 3 bugs in public open-source repos as portfolio samples (with permission, via their issue trackers). Days 3–4, message 20 indie makers or small teams offering one free bug diagnosis. Days 5–7, convert to paid — the goal is 1 paying client or 5 warm leads. If nobody replies after 20 genuine outreaches, drop it: the setup cost you nothing but time.

Pitfalls and limits

• Do not self-host on a laptop. A 1.05-trillion-parameter model needs data-center-scale GPUs; the download is for server teams. Use the API or Vibe Code.
• Vendor benchmarks are vendor benchmarks. Every score above is Mistral’s own. Treat “leads open models” as a claim to test on your own bugs, not a fact.
• The license catch matters if you resell access. Selling fixes is fine; selling other people API access to Large 4 itself needs Mistral’s commercial license.
• It can still write wrong code confidently. Always read the fix, run the project’s tests, and never paste client secrets or private keys into any hosted chat.
• Preview vs. open weights, don’t mix them up. Articles from October 6 describe the preview; the weights, the improved scores, and the $0.70 price all date from October 29.

Sources

• Mistral AI official post, October 29, 2026: “Mistral Large 4 (Le Chonk) is now open weight” — weights release, reasoning modes, benchmarks, $0.70/$0.07/$2.10 pricing, license terms, and availability channels (Studio, API, OpenRouter, Hugging Face, Microsoft Foundry, NVIDIA NIM, Vibe Code).
• Mistral Large 4 model card (docs.mistral.ai) — preview specifications: 1.05T parameters, 1M context, native multimodality.
• Related on this site: Voxtral 3 Guide (Mistral’s phone-answering voice AI) and Timestral Guide (Mistral’s no-training forecasting AI) — same vendor, different jobs.