Mistral Large 3
Mistral AIMistral AI's first mixture-of-experts model since Mixtral and the current open-weights flagship, released December 2, 2025 under Apache 2.0. 41B active / 675B total parameters, 256K context, native multimodal support (text + images). Trained from scratch on 3000 H200 GPUs. Debuted at #2 on the LMArena OSS non-reasoning leaderboard. Combined with Ministral 3 (smaller dense variants) and Mistral Vibe CLI in the broader Mistral 3 stack. Runs at the compute cost of a ~41B dense model while accessing the capacity of a 675B one, Europe's flagship open-weights frontier model and the direct successor to the retired Mistral Large 2 (March 2025).
Mistral Large 3 (mistral-large-3-25-12), announced 2 Dec 2025. Sparse MoE, 41B active / 675B total params. Base + instruct released; reasoning variant announced but not shipped as of mid-2026. Consolidated on 2026-07-29 from mistral-large-2; openness follows mistral-large-3, the release that currently governs. See docs/reference/identity.md. The LMArena placement is the launch-day figure from Mistral's own post, not a current leaderboard read. Adoption is low because the flagship is consumed through the API rather than downloaded, and the map bands on the declared checkpoints. Verified 2026-08-13 via the Mistral 3 launch post and the Mistral-Large-3-675B-Base-2512 HF model card.
Openness
3 high confidence- weights
- open(Apache-2.0)
- data
- closed
- code
- closed
- license
- Apache-2.0(OSI)
Apache-2.0 weights with closed data and no training code, so 3. Mistral relicensed between releases: Large 2 shipped under the Mistral Research License, which permits research and non-commercial use only and scored 2, and Large 3 is Apache-2.0. Large 3 is the current release, so the score follows the relicensing rather than the family's worst historical terms.
- https://huggingface.co/api/models/mistralai/Mistral-Large-3-675B-Base-2512 recorded 2026-08-13
card license `apache-2.0`, `gated: false`, `private: false`; the base checkpoint, distinct from the -Instruct-2512 sibling.
- https://mistral.ai/news/mistral-3/ recorded 2026-08-13
"We release both the base and instruction fine-tuned versions of Mistral Large 3 under the Apache 2.0 license"; "trained from scratch on 3000 of NVIDIA's H200 GPUs". The post names no corpus and links no training pipeline.
- https://docs.mistral.ai/models/mistral-large-3-25-12 recorded 2026-08-15
Mistral's model reference for mistral-large-3-25-12, describing it as open-weight with a 256k context. Corroborates the weights reading against the vendor's own docs.
- https://azure.microsoft.com/en-us/blog/introducing-mistral-large-3-in-microsoft-foundry-open-capable-and-ready-for-production-workloads/ recorded 2026-08-15
Microsoft's launch post for Mistral Large 3 in Foundry, describing it as an open-weight model available for production workloads. A distribution partner's description; it establishes no dimension the vendor's own sources do not.
Adoption
1 medium confidence5,971 downloads in the trailing 30 days across the declared artifacts (mistralai/Mistral-Large-Instruct-2407 4,714; mistralai/Mistral-Large-3-675B-Base-2512 80; mistralai/Mistral-Large-3-675B-Instruct-2512 1,177), which bands at level 1 (<10K) on the model adoption scale. The flagship is mostly consumed through Mistral's API rather than downloaded, so this understates real use - but the map bands on the artifacts a product declares, and this is the label those support.
- https://huggingface.co/api/models/mistralai/Mistral-Large-Instruct-2407 recorded 2026-08-12
4,714 downloads in the trailing 30 days for mistralai/Mistral-Large-Instruct-2407
- https://huggingface.co/api/models/mistralai/Mistral-Large-3-675B-Base-2512 recorded 2026-08-12
80 downloads in the trailing 30 days for mistralai/Mistral-Large-3-675B-Base-2512
- https://huggingface.co/api/models/mistralai/Mistral-Large-3-675B-Instruct-2512 recorded 2026-08-12
1,177 downloads in the trailing 30 days for mistralai/Mistral-Large-3-675B-Instruct-2512
Capability
4 high confidenceDebuts #2 in LMArena OSS non-reasoning category (#6 among all OSS models). Per-benchmark MMLU-Pro/GPQA/SWE-bench not given in the launch post. The ranking is a launch-day placement rather than a current leaderboard read, which is the limit of what this source can support and the reason the band is not a 5.
- https://mistral.ai/news/mistral-3/ recorded 2026-08-13
"Mistral Large 3 debuts at #2 in the OSS non-reasoning models category (#6 amongst OSS models overall) on the LMArena leaderboard"
Verified 2026-08-12