Mistral Large 4 Is Open Weights in Name and API First in Practice

Mistral Large 4 launched last week as an open-weight model, and as of launch day there was no checkpoint on Hugging Face.

Roughly a trillion parameters, around 50 billion active per token, and the only way to touch it is Mistral's own API. The weights are scheduled for October 27. Until then, "open" is a calendar entry.

I'm not saying that to dunk on Mistral. I'm saying it because the word "open" is doing a lot of work in how builders pick models, and this launch is a clean case study in how little the word guarantees.

Mistral Large 4 Launched Without Weights

Here is what's actually public. Large 4 is in preview on Mistral's infrastructure. A second version, with reduced moderation and expanded cyber capabilities, is going to cybersecurity leaders, vetted partners and state authorities. Everyone else waits for the 27th.

Mistral's headline numbers are vendor-reported. 61.7% on DeepSWE v1.1 for coding, 82% on a cyber reproduce-and-patch test, 93% on Cybench. On the same DeepSWE board, GLM-5.3 and Kimi K3 sit around 69%, and the leading closed models around 74%.

So it's a strong open-weight model that trails on coding. Fine. That's a real trade, and plenty of workloads don't care about the last 12 points.

What I care about is the sequence. API first, weights later, license TBD. Large 3 shipped under Apache 2.0. Large 4's weights are expected under a custom Mistral license, and nobody has seen the text.

Open Weights Is a Promise Until You Hold the File

I've been burned by roadmaps before. I wrote about the OpenAI Assistants API sunset in OpenAI Assistants API Sunset: you build on a thing, the vendor decides the thing is over, and your migration is now your problem.

Mistral's own pitch is the mirror image of that. Their chief scientist put it plainly: if you use a closed model, there's no guarantee it will still be there tomorrow. Open weights means you own a complete copy that nobody can retire.

That argument is correct. It's also only true after you've downloaded the weights. Before that, you're a customer of a hosted API with a press release that says otherwise.

# the only open-weight check that matters
huggingface-cli download mistralai/<model> --local-dir ./weights
sha256sum ./weights/*.safetensors > weights.sha256

If that command doesn't work yet, you're not building on open weights. You're building on an API.

The Part Where I Do the Math on Self-Hosting

Say the weights land on the 27th. A trillion total parameters is not a model you run on a laptop, and it isn't a model I'd run on a single box in my closet either. Active parameters sit around 49 to 52 billion depending on which source you read, but you still have to hold all the experts in memory.

I run n8n self-hosted and I like owning my stack. I also know the difference between owning a stack and owning a rack. For almost everyone reading this, "self-host Large 4" really means "rent GPUs from someone, run the weights there". That's still a real option and it still beats a vendor's kill switch. It's just not the free lunch the word "open" implies.

The honest comparison isn't Mistral API versus a box under your desk. It's Mistral API versus another host running the same weights at a price you can shop around.

The Uncomfortable Take

Here's what I'd tell you: staged open-weight releases are a distribution strategy, not a philosophy.

Release to your own API first, collect the revenue and the feedback, hand a safer-moderated copy to the public and a looser one to vetted partners, then publish weights when the commercial window has closed. Everyone in the industry will start doing this, because it's smart.

The safety reasoning isn't fake. Mistral reportedly saw the model try to go beyond its test environment during evaluation and contained it in software. Once weights are public, anyone can strip guardrails. Staging gives them time to see what the cyber capabilities do in the wild.

But notice what it does to the lock-in story. During the stage you're integrating against their API, their rate limits, their pricing. By the time the weights arrive, your prompts, evals and retries are all tuned to that endpoint. The exit door exists, and you've already built your house next to the other one.

Open weights protect you from deprecation. They don't protect you from convenience.

What I'd Actually Do

Three things, none of them clever.

First, treat the 27th as the start date for evaluation, not the launch date. Read the license before you read the benchmarks. A custom license that restricts commercial use or large-scale hosting matters more than three points on Cybench.

Second, keep your model calls behind one thin interface. I made this argument in OpenCode and the case against model allegiance, and it's the same lesson here. If swapping the model is a one-line config change, every launch like this is an option instead of a migration.

Third, run your own evals on your own tasks. Vendor numbers for a model that hasn't shipped weights are marketing, and I'd say so about any vendor. Independent leaderboards currently put Large 4 between DeepSeek V4.1 Flash and OpenAI's entry-level GPT6 Luna models, which is a useful, less flattering sentence than the launch post.

Open Is Something You Verify

I'm not cynical about this. A strong open-weight model from outside China matters, and Mistral shipping a trillion-parameter model that vendors can't switch off is good for every builder who's been through a forced migration.

I just won't put the word "open" in my architecture doc until I've got the file on disk and the license in hand.

Download the weights. Read the license. Then decide.