benchr
Most of the models that top the leaderboards live behind an API you can rent but never own. Mistral Large 3 is the opposite bet. It's the largest open-weight model any major lab has shipped, a 675B-parameter mixture of experts that Mistral released under the Apache-2.0 license and published on Hugging Face. Mistral's own framing for the launch is blunt: "the future of AI is open." You can download the weights, run them on your own hardware, and ship commercially with no licensing fee. For a model at this scale, that's the headline, and it's true.
It's worth being precise about the wider story here, because the press writes a lot of it. Mistral is a Paris-based lab, and it's widely cast as the leading independent European open-weight provider, which is part of why EU buyers weighing data-sovereignty rules keep coming back to it. That positioning is real in the market, but it's market positioning, not an official product claim. Mistral's Mistral 3 announcement doesn't lead on European sovereignty at all. It leads on openness. The concrete, verifiable thing a sovereignty-minded buyer gets is the Apache-2.0 license, which means the weights can run entirely inside your own walls. Everything past that is reputation, and reputation isn't a spec.
Which models this covers
The slug just says "mistral," so a word on what's current. This review is about the Mistral 3 family announced December 2, 2025, not the older Large 2 line. The flagship open-weight model is Mistral Large 3, API id mistral-large-2512, version 25.12. If you're provisioning against an older Large 2 model id, you're on the previous generation.
There's a genuine wrinkle in what counts as "the flagship," and it's worth stating plainly rather than papering over. Mistral describes Large 3 as its state-of-the-art, open-weight, general-purpose multimodal model, the largest and most capable open release. But it also calls Mistral Medium 3.5 (version 26.04) its frontier-class model for agentic and coding use, and Medium 3.5 is the default model in Le Chat and Mistral Vibe. So there are two answers depending on what you mean: Large 3 is the headline open-weight flagship, and Medium 3.5 is the agentic, API-priced one that most chat users touch. The honest read is to hold both in your head. Below is the whole current lineup in one view.
| Model | Price (in / out, per 1M) | License | Context |
|---|---|---|---|
| Mistral Large 3 (v25.12) — open-weight flagship | $0.50 / $1.50 | Apache-2.0 (open weights) | 256K |
| Mistral Medium 3.5 (v26.04) — agentic / coding, in preview | $1.50 / $7.50 | Open weights + API; exact open license unconfirmed | 256K (third-party figure) |
| Mistral Small 4 (v26.03) — small open model | $0.10 / $0.30 | Open weights | Not specified on the fetched docs page |
A couple of honest flags on that table. The 256K context for Medium 3.5 comes from third-party and Hugging Face listings, not from a figure Mistral's docs displayed on the page checked, so treat it as reported rather than confirmed. The exact open-weight license for Medium 3.5 couldn't be pinned to an official Mistral or Hugging Face page either, so it's left as unconfirmed rather than guessed at. And Medium 3.5 is in public preview, which means its specs and pricing can shift before it stabilizes. None of that changes the shape of the lineup, but it should change how hard you lean on the preview model's numbers.
No benchmark table, so don't buy on benchmarks
Here's the thing that keeps this review from leaning on scores: Mistral's docs model pages don't display a benchmark table for these models. The figures that float around come from Mistral's own announcements, where they're self-reported, or from third-party leaderboards. There's no neutral, audited table to quote, so this page doesn't quote one, and you shouldn't either when you're deciding. A vendor-reported number tells you a ceiling under the vendor's own harness, not how the model behaves on your tasks. For why a single headline figure is a weak basis for a buying call in the first place, benchr's piece on why the benchmarks stopped telling you anything is the right background.
That pushes the decision onto things you can verify: the license, the price, the openness, the context window, and the breadth of inputs and languages. Those are the columns that hold up. So that's where this review argues.
What you're buying
Start with scale and shape. Large 3 is a mixture of experts with around 675B total parameters and roughly 41B active per token, which is how a model this big can serve at a cost that doesn't track its full parameter count. It's natively multimodal, with a vision encoder fused in at around 2.5B parameters, so image understanding isn't bolted on after the fact. And it's multilingual across more than 40 languages, which matters if your users aren't all writing in English. For where Mistral's image handling sits against the rest of the field, benchr's multimodal capability ranking is the cross-shop.
Then there's the context. All three models in the lineup carry a 256K window on the figures available, and on Large 3 that 256K is confirmed on Mistral's own docs. A window that size holds a real codebase or a large document set in a single prompt instead of stitching it together with retrieval. The usual caution applies: accepting a 256K context isn't the same as using all of it accurately, and recall degrades across long inputs for every model, so benchmark retrieval on your own long prompts. benchr's roundup of context windows compared sets the baseline to measure Mistral against, and the look at how big-context claims get marketed is the reality check on what a large number buys you.
What it costs to run
Two ways to pay, and they tell different stories. The first is the Apache-2.0 license, which costs nothing. Download Large 3, run it on your own hardware, ship commercially, no fee. The catch is that a 675B-parameter mixture of experts is not a model you stand up on a spare GPU. Self-hosting it means a serious multi-GPU box and the engineering to keep it serving, which only pays off at high, steady volume or when data legally can't leave your network. benchr's guide to running models on your own machine walks through where that line sits, and for weights this size it sits high.
The second is the hosted API, and this is where most people should land. Large 3 bills $0.50 per million input tokens and $1.50 per million output on Mistral's official pricing page. That's cheap against closed frontier output that runs into dollars per million. Medium 3.5 is pricier at $1.50 in and $7.50 out, the premium you pay for the agentic and coding tuning, while Small 4 drops to $0.10 in and $0.30 out for high-volume simpler work. To see which tier pays off at your real call volume rather than at the sticker rate, benchr's price-per-use-case breakdown is the tool, since a low per-token price either dominates the bill or barely shows up depending on how often you hit the model.
Where it sits against the field
The open-weight tier is crowded now, and Large 3's claim is breadth at the very top of the size range: it's the biggest open model from a major lab, multimodal and multilingual, under a permissive license. The natural cross-shop is the other large open mixture-of-experts releases. DeepSeek competes hard on raw coding with its own giant MoE weights, and benchr's DeepSeek-V4 review lays out that trade; Meta's open flagship is the other obvious comparison, covered in benchr's Llama 4 review. Against those, Mistral's differentiators are the native multimodality, the 40-plus-language reach, and the clean Apache-2.0 terms.
Where to be cautious is exactly the place the benchmarks would normally settle: capability ranking. Without an official audited table, the honest answer is that you should treat Large 3 as frontier-scale on paper and prove it on your workload before betting production on it. The openness is verifiable today. The relative quality against DeepSeek, Llama, and the closed frontier is something you confirm on your own tasks.
The verdict
Mistral Large 3 is the strongest case going that frontier-scale capability and open access aren't a contradiction. The largest open-weight model from a major lab, Apache-2.0, multimodal, multilingual, 256K context, and a hosted API cheap enough that the open weights are insurance rather than the only way in. The 4.1 here reflects that the value and the openness are real, held short of the top by the one thing missing: Mistral publishes no official benchmark table for these models, so you're buying on license, price, and reach, then proving capability yourself.
Go with Mistral Large 3 if you want the biggest open model you can either self-host under Apache-2.0 or call cheaply through the API with no lock-in, and your priorities are openness, multimodality, languages, and a long context. Reach for Mistral Medium 3.5 when the work is agentic or coding-heavy and you want Mistral's purpose-built model for it, accepting the higher per-token cost and the preview tag. Drop to Mistral Small 4 for high-volume, simpler tasks at a tenth the input price. Skip the family only if you need an independently audited capability ranking before you'll commit, because that's the one thing the official sources don't give you yet.
Frequently asked
Is Mistral Large 3 open source?
The weights are open. Mistral Large 3 is released under the Apache-2.0 license, as stated in Mistral's own Mistral 3 announcement, and the model is published on Hugging Face. Apache-2.0 is a permissive license with no restriction on commercial use, so you can download the weights and self-host at no licensing cost. What is not free is the hardware to run a 675B-parameter mixture-of-experts model, or the hosted API, which is billed per token.
What is the difference between Mistral Large 3 and Mistral Medium 3.5?
Large 3 (version 25.12, API id mistral-large-2512) is the open-weight flagship: a 675B-parameter mixture of experts with about 41B active, released under Apache-2.0, natively multimodal and multilingual with a 256K context. Mistral Medium 3.5 (version 26.04) is the smaller, denser model Mistral describes as frontier-class for agentic and coding work, and it is the default model in Le Chat and Mistral Vibe. Pick Large 3 when you want the biggest open model you can host or call cheaply; pick Medium 3.5 for agentic and coding workloads through the API. Note that Medium 3.5 is in public preview, so its specs may change.
How much does the Mistral API cost?
From Mistral's official pricing page, per million tokens: Mistral Large 3 is $0.50 input and $1.50 output; Mistral Medium 3.5 is $1.50 input and $7.50 output; Mistral Small 4 is $0.10 input and $0.30 output. The small Ministral models run cheaper still, from $0.10 to $0.20 per million on both sides. Le Chat also offers a free tier if you just want to use the models without an API key. Verify the live rate before you budget, since Medium 3.5 is a preview and prices can move.
Is Mistral a European AI company, and does that matter for data sovereignty?
Mistral is a Paris-based lab, and it is widely described in the press as the leading independent European open-weight provider, which makes it an appealing option for EU buyers weighing data-sovereignty requirements. Treat that as market and press positioning rather than an official product claim: Mistral's own Mistral 3 announcement does not frame the launch around European sovereignty. It leads instead with the line that the future of AI is open. The concrete, verifiable advantage for sovereignty-minded buyers is the Apache-2.0 license itself, which lets you take the open weights and run them entirely on your own infrastructure.
Which Mistral model should I pick?
Go with Mistral Large 3 if you want frontier-scale capability you can either host yourself under Apache-2.0 or call through a cheap hosted API at $0.50 input and $1.50 output. Reach for Mistral Medium 3.5 when your work is agentic or coding-heavy and you want Mistral's purpose-built model for that, accepting that it is pricier per token and still in preview. Drop to Mistral Small 4 at $0.10 input and $0.30 output for high-volume, simpler tasks where you do not need the flagship. Across all three, the long 256K context and the multimodal, multilingual reach come standard.
Changelog
- May 30, 2026 — Originally published. Lineup, license, context window, and per-token pricing verified against Mistral's official Mistral 3 announcement (mistral.ai/news/mistral-3), the docs model overview and the Mistral Large 3 model page (docs.mistral.ai), the official pricing page (mistral.ai/pricing), and the Hugging Face model card. No official benchmark table is published for these models, so no scores are cited. The European-sovereignty framing is attributed to press positioning, not to Mistral's own announcement; Mistral Medium 3.5 is noted as a public preview, and its 256K context and open-weight license are flagged as reported rather than officially confirmed.
References
- Mistral AI, "Mistral 3," mistral.ai/news/mistral-3, December 2, 2025.
- Mistral AI, "Models overview," docs.mistral.ai/models/overview, accessed May 2026.
- Mistral AI, "Mistral Large 3 (25.12) model page," docs.mistral.ai/models/mistral-large-3-25-12, accessed May 2026.
- Mistral AI, "Pricing," mistral.ai/pricing, accessed May 2026.
- Mistral AI, "Vibe remote agents and Mistral Medium 3.5," mistral.ai/news/vibe-remote-agents-mistral-medium-3-5, accessed May 2026.
- Mistral AI, "Mistral-Large-3-675B-Instruct-2512 model card," huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512, accessed May 2026.