Mistral Large 4: Pricing, Specs, Benchmarks & How to Use It

What if you could rent a trillion-parameter AI model for about a third of what the top closed models charge? That’s the pitch behind Mistral Large 4, which went into public preview on October 6, 2026. It’s a mixture-of-experts model with roughly 1 trillion total parameters, about 49 billion active per token, image and text input, and an API you can call today, with open weights promised by October 31.

So is it the new king of AI? Not quite. Independent testing puts it well behind the best closed models from Anthropic, OpenAI and Google. But it’s Mistral’s biggest model ever, it’s unusually strong at cybersecurity, and it’ll soon be something you can host yourself. Here’s what it costs, what’s inside it, how it scores, and how you can try it.

Detail	Mistral Large 4
Developer	Mistral AI (France)
Released	October 6, 2026 (public preview, version v26.10)
API model ID	mistral-large-4
Total parameters	About 1.05 trillion (mixture of experts)
Active parameters	49B per token (52B counting embeddings and output layers)
Context window	1M tokens per Mistral’s docs; 524K measured on the preview by Artificial Analysis
Input / output	Text and images in, text out
List price	$1.36 input / $4.18 output per 1M tokens
Price shown in docs now	$0.68 input / $2.09 output / $0.07 cached input
Open weights	Expected October 31, 2026 on Hugging Face
Languages	160+, including all official EU languages
Nickname	Le Chonk

What is Mistral Large 4, and why should you care?

Mistral Large 4 (Mistral shortens it to ML4) is the Paris lab’s new flagship. It blends a fast instruct mode and a step-by-step reasoning mode in a single model, so you don’t have to pick between separate variants. According to Mistral’s official announcement, it was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacenters.

Why does that matter to you? Because it’s the first big product from Mistral’s €3 billion Series D, which the company calls the largest equity round ever raised by a European tech firm. Tech.eu reports the round closed in September at a valuation of roughly $24 billion. ML4 is the first milestone on the roadmap that money pays for.

The funding race isn’t just a European story, either. China’s DeepSeek is reportedly raising far bigger sums for its own open models, and those Chinese labs are exactly who Mistral is now chasing. The key takeaway here is simple: Europe finally has a trillion-parameter, soon-to-be-open model of its own.

Mistral Large 4 specs, explained in plain English

Let’s break it down. A mixture-of-experts (MoE) model is a big pool of smaller sub-networks, called experts. For each token, a router wakes up only the experts it needs, so you get the knowledge of a huge model with the running cost of a much smaller one. ML4 holds about 1.05 trillion parameters but uses only a slice of them at a time.

You may have seen two different numbers for that slice: 49 billion and 52 billion. Both are right. The Hugging Face page says 49 billion parameters are active per token, or 52 billion if you count the embedding and output layers. That’s why the repo is named Mistral-Large-4.0-1T05-A52B, meaning 1.05 trillion total and 52 billion active. Mistral’s docs also list a separate 1.6-billion-parameter vision encoder for images.

Context length is less settled. Mistral’s model docs list a 1-million-token window, and so does Ollama. Artificial Analysis, though, measured 524,000 tokens on the preview endpoint, and MindStudio’s write-up describes it as roughly 500,000 tokens. If your workflow depends on feeding in huge codebases or document sets, test the real limit before you commit.

Bottom line: big brain, modest per-token compute, and a context window you should verify for yourself.

How much does Mistral Large 4 cost?

Mistral’s launch post lists $1.36 per million input tokens and $4.18 per million output tokens. At the time of writing, its docs and Ollama’s listing show half that, $0.68 input and $2.09 output, with cached input at $0.07, displayed as a discount with the original prices struck through. Mistral doesn’t say how long that lower rate lasts, so plan your budget around the list price.

Model	Input (per 1M tokens)	Output (per 1M tokens)	Open weights?
Mistral Large 4 (list)	$1.36	$4.18	Due Oct 31
Mistral Large 4 (current docs price)	$0.68	$2.09	Due Oct 31
Mistral Small 4	$0.15	$0.60	Yes, Apache 2.0
Claude Sonnet 5.5	$2	$10	No
Claude Opus 5.5	$4	$20	No

Claude figures come from Anthropic’s pricing page; Mistral Small 4 pricing comes from Mistral’s Small 4 launch post.

Let’s put numbers on it. Say your app burns through 10 million input tokens and 2 million output tokens a month. At list price, ML4 costs about $21.96. The same job runs about $40 on Claude Sonnet 5.5 and about $80 on Claude Opus 5.5.

Here’s the catch. Artificial Analysis found ML4 is chatty: it produced 200 million output tokens while running the firm’s Intelligence Index, against a median of 81 million for comparable models. More words mean more output tokens on your bill, so your real savings depend on how much it thinks out loud for your tasks.

Benchmarks: where Mistral Large 4 shines and where it trails

Mistral’s own charts focus on coding, agents and security. These are the company’s numbers, so treat them as a starting point, not the final word:

• Coding: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.
• Agents: 59.9% on AutomationBench, which covers 657 business workflows.
• Cybersecurity: 93% of 40 Cybench challenges and 82% on a reproduce-and-patch test, which Mistral says is the highest of any model.
• Vision: 42% on a dense visual grounding test, edging GPT-6-Astra’s 41%.
• Human coding review: 3.74 in a blind Surge AI evaluation, second of five models, behind Claude Opus 5 at 4.22.

Now the independent view. Artificial Analysis gives the reasoning preview a score of 38 on its Intelligence Index, ranked 64th of 225 models it tracks, with fast output of about 116 tokens per second and a 1.46-second wait for the first token. The Register notes that this places ML4 between DeepSeek V4.1 Flash and OpenAI’s GPT-6 Luna, a long way behind the flagship closed models.

Against other open-weight models, the picture is mixed. Here’s how MindStudio’s comparison lines up on shared tests (scores can vary with the coding harness each lab used):

Benchmark	Mistral Large 4	GLM 5.3	Kimi K3
DeepSWE 1.1	62	Ranked higher (score not given)	68
Terminal-Bench	28.3	40	Not listed
AutomationBench	59.9	62.2	58.3

You see, ML4 isn’t the best open model in the world. Chinese labs still lead on general coding and agent tasks. Where it does pull ahead is cybersecurity, where MindStudio says it ties for first on the Artificial Analysis Cyber Index with GLM 5.3 Flash. Mistral frames that as the headline:

“ML4 pairs top-tier cyber performance with open weights and self-deployment.”

That line, from Mistral’s launch post, matters because security teams often hit refusals from closed models when they test attacks. Cyber is now a frontier battleground in its own right; Google, for example, has limited Gemini 4 Argon to vetted cyber defenders. The key takeaway: pick ML4 for security and enterprise workloads, not for raw coding supremacy.

How to try Mistral Large 4 today

The weights aren’t out yet, so for now you reach ML4 through an API. MindStudio reports it isn’t in Mistral’s consumer chat app at launch, so plan on a developer route.

1. Open a Mistral Studio account and create an API key

Sign in to Mistral Studio, Mistral’s developer console, and generate a key. Mistral’s pricing page notes the free plan includes Studio access, though you should expect usage limits and billing for heavier use.

2. Call the mistral-large-4 model ID

Point your code at the model ID mistral-large-4. The docs list support for the Chat Completions, Conversations and Batch endpoints, plus function calling, structured outputs, document Q&A and built-in tools. If you already use another Mistral model, swapping the model name is often all it takes.

3. Use Ollama’s cloud tag if you live in the terminal

Ollama lists a mistral-large-4:cloud tag, so ollama run mistral-large-4:cloud gets you chatting fast. It also offers launch commands that plug ML4 into coding agents such as Claude Code and OpenCode. This runs on Ollama’s servers, not your machine.

4. Watch the reasoning tokens in coding tools

MindStudio reports runaway chain-of-thought in OpenCode that can fill the context window without a clear error. Set output limits, log token usage, and start with small tasks before you let it loose on a big repo.

Can you run Mistral Large 4 on your own hardware?

Short answer: not on a gaming PC or a laptop. Do the math. Roughly 1.05 trillion parameters at 8 bits each is about 1 terabyte of weights, and even an aggressive 4-bit version would need around 525 GB before you add memory for the context. No consumer GPU comes close.

Mistral hasn’t published hardware requirements. The Register’s reporter estimates it should run reasonably on 8-GPU servers like Nvidia’s HGX B300 or AMD’s MI355X systems. That’s data-center gear, which is why self-hosting makes sense for companies and governments, not hobbyists.

If you want a Mistral model on your own box, look at Mistral Small 4 instead. It has 119 billion total parameters, only 6 billion active, a 256K context window, and an Apache 2.0 license. It’s still big, but it’s within reach of a well-equipped workstation.

Who should use Mistral Large 4, and who should wait?

ML4 isn’t for everyone. Here’s how to tell if it fits you:

• European and regulated teams: Mistral runs a European deployment end to end and promises private-cloud and on-premise options, which can simplify data-residency conversations. Samon’s AI Brief argues this gives regulated buyers real procurement leverage.
• Security teams: strong cyber scores plus self-hosting means you can test attacks without a vendor refusing the request.
• High-volume, cost-sensitive apps: list prices well below Claude’s make it worth a side-by-side trial on your own data.
• Wait if you need peak reasoning or coding: the top closed models still score far higher on independent tests.
• Wait if you’re a solo developer on a deadline: early integrations with coding agents are rough, and the preview may change.

That said, the open-weights release changes the math for anyone who wants control. Once the files land, you can fine-tune, host it where you like, and stop worrying about a vendor changing terms. If you’re new to AI-assisted coding, our guide to vibe coding explains how these models slot into your workflow.

Mistral Large 4 is worth a test drive: here’s your next move

Le Chonk won’t knock the leading closed models off their perch, and Mistral isn’t pretending otherwise. What it offers is a rare mix: near-frontier scale, strong security skills, European hosting, and open weights in a few weeks. That’s a meaningful choice for a lot of teams.

So grab an API key, run your own prompts through it while preview pricing is low, and compare the bill with what you pay now. Then mark October 31 on your calendar. If the weights arrive on time, you’ll have one of the most capable models you can own outright.

Frequently asked questions