Mistral Large 2 vs Large 3: Very Different Models, One Name

"Mistral Large" is not one model. It is a product name that has covered several releases, two of them still common in catalogues, with different architectures, different licences and prices that differ by a factor of four. If your configuration says mistral-large and nothing else, the first job is to find out which one you are actually getting.

This guide separates the two, explains what changed with Mistral Large 3, and covers the question Mistral's own deprecation table raises: whether you should be on Large at all. A third generation has also appeared: Mistral's documentation lists Mistral Large 4 as a public preview from 6 October 2026, an open-weight MoE with 1.05T total and 52B active parameters and a 1M-token window. It is too new and still in preview, so this guide does not cover it; check its licence and status before planning around it.

One name, two different models

Mistral Large 2 arrived on 24 July 2024 as mistral-large-2407. It was a dense 123B model with a 128K context window, designed for single-node inference, and Mistral claimed support for more than 80 programming languages and 84.0% on MMLU for the pretrained version. The weights were published under the Mistral Research License: free for research and non-commercial use, with a separate commercial licence needed to self-deploy it in a product.

Mistral Large 3 arrived on 2 December 2025 as mistral-large-2512. It is a sparse mixture-of-experts model with 41B active and 675B total parameters, a 256K context window, image input, and Apache 2.0 weights. Mistral describes it as its first MoE since the Mixtral series, trained from scratch on 3,000 NVIDIA H200 GPUs.

Those are not a point release apart. They differ in how they run, what you may do with the weights, and what they cost. OpenRouter's catalogue, for instance, still describes its unversioned mistral-large entry as Large 2 (2407), listed at $2 input and $6 output per million tokens at the time of writing. Mistral's own documentation lists Large 2.0 as retired since March 2025 and the later 2411 build as retired on 31 May 2026. Whatever an unversioned alias resolves to today, confirm it rather than assume it. Pinning model versions covers why aliases break production quietly.

Large 3: sparse, Apache 2.0, and cheaper

The move to mixture-of-experts is the change that drives everything else. Only around 41B parameters run per token, so compute per token is a fraction of what the 675B total implies. The model card is more precise: a 673B language model with 39B active, plus a 2.5B vision encoder.

That shows in the price. Mistral lists Large 3 at $0.50 per million input tokens and $1.50 per million output, and OpenRouter matches it at the time of writing. A :batch variant is the same model on an asynchronous endpoint at half that, $0.25 and $0.75. Against the $2 and $6 listed for Large 2, Large 3 is cheaper on both sides by a factor of four.

The licence change matters as much as the price. Apache 2.0 means you can fine-tune, redistribute and serve it commercially without asking anyone. That is a different proposition from Large 2's research licence, and from the modified MIT terms Mistral later used for Medium 3.5. Open-weight model licences compared covers where those differences start to bite.

At launch Mistral reported Large 3 at second among open-source non-reasoning models and sixth among all open-source models on the LMArena leaderboard. The model card's one numeric score in text is 67.17 on GPQA Diamond. A reasoning version was described as coming soon at release; check its current status before planning around it.

What Mistral says it is not good at

The model card is unusually candid about limitations, and they are worth taking at face value:

• It is not a dedicated reasoning model. Models built for extended reasoning can beat it on strict maths and logic tasks.
• Vision is secondary. It can trail vision-first models on multimodal work.
• Deployment is complex because of its size and architecture.

The recommended uses follow from that: long-document understanding, a general assistant, agentic and tool-use work, enterprise knowledge tasks and general coding help. It is a broad generalist, not a specialist at anything.

Self-hosting a 675B model

Apache 2.0 weights are only useful if you can run them. The model card gives two practical routes: the FP8 checkpoint on a single node of B200s or H200s, and an NVFP4 checkpoint on a single node of H100s or A100s. A BF16 version also exists.

A full node of B200s or H100s is a serious piece of hardware, but it is one node, not a cluster. The total parameter count sets your memory requirement; the 41B active sets your throughput. If that distinction is new, dense vs mixture-of-experts covers it. Mistral recommends a temperature below 0.1 for production use.

Large 3 or Medium 3.5

Here is the awkward part. Mistral's deprecation table names Large 3 as the replacement for Large 2.0, but names Mistral Medium 3.5 as the replacement for Large 2.1. Medium 3.5, released at the end of April 2026, is a dense 128B model with per-request reasoning effort and a vendor-reported 77.6% on SWE-bench Verified.

The two are not interchangeable:

• Price — Large 3 lists at $0.50/$1.50; Medium 3.5 at $1.50/$7.50. On output-heavy work Large 3 is five times cheaper.
• Licence — Large 3 is Apache 2.0. Medium 3.5 is modified MIT, with a revenue threshold for large companies.
• Reasoning — Medium 3.5 has a reasoning mode you switch on per request. Large 3, by Mistral's own account, is not a reasoning model.
• Hosting — Large 3 needs a full node of data-centre GPUs. Mistral says Medium 3.5 can run on as few as four GPUs.

A reasonable default: Large 3 for high-volume general work where cost and a permissive licence matter, Medium 3.5 for coding agents and tasks that benefit from reasoning. The Mistral Medium 3.5 guide covers that model in detail.

A checklist before you switch

1. Replace any unversioned mistral-large reference with an explicit dated id such as mistral-large-2512.
2. If you self-host Large 2 under a commercial agreement, check whether moving to Apache 2.0 weights simplifies that contract.
3. Run twenty real tasks through Large 3 and Medium 3.5 on the same harness. Record pass rate and output tokens, then compute cost per completed task.
4. If your work is mostly retrieval and summarisation, start with Large 3. If it is multi-step code changes, start with Medium 3.5.