Mistral Large 3
Mistral Large 3 is Mistral AI's biggest open-weight model: a 675-billion-parameter mixture-of-experts model released on 2 December 2025 under the Apache 2.0 licence. Only 41 billion parameters are active for each token, which keeps it relatively cheap to run. It reads text and images, handles a 256,000-token context and supports dozens of languages. On Mistral's API it costs $0.50 per million input tokens and $1.50 per million output tokens.
It is a "non-reasoning" model, meaning it answers directly rather than thinking step by step first. At launch Mistral said it ranked #2 among open non-reasoning models on LMArena. By September 2026 it has fallen behind: it scores 9 on the Artificial Analysis Intelligence Index, below Mistral's own Medium 3.5. It remains a cheap, fully open, multilingual workhorse that runs on a single 8-GPU server.
Mistral Large 3 is a cheap, fully open, multilingual model that is easy to deploy, but it is no longer competitive on hard reasoning or coding.
What it offers:
- A true open licence. Apache 2.0, with no revenue caps or attribution rules, unlike Medium 3.5, MiniMax M3 or GLM-5.3.
- Low price. $0.50/$1.50 per million tokens on Mistral's API, and wide availability on Bedrock, Azure and other clouds.
- One-server hosting. The FP8 version fits on a single 8x H200 node, and NVFP4 on one H100 or A100 node.
- Languages and images. Good multilingual chat and image understanding.
The weaknesses: it does not reason step by step, and independent testing puts it well behind current models. It scores 9 on the Artificial Analysis Intelligence Index, against 14 for Medium 3.5 and 45 for GLM-5.3. Artificial Analysis also found it verbose.
Pick it if you need a large open model under Apache 2.0 for multilingual chat, summarising, translation or document work, especially inside the EU.
Skip it if you need coding, maths or agent work. Choose Medium 3.5, GLM-5.3 or a closed model such as Claude Opus 5.5.
Score breakdown
Best for
- Multilingual chatbots and translation
- EU-hosted or on-premises deployments needing a permissive licence
- Summarising and extracting from long documents and images
- Fine-tuning a large open base model
What we like
- Apache 2.0: no revenue caps or attribution requirements
- Cheap at $0.50 in / $1.50 out per 1M tokens
- Fits on a single 8-GPU node in FP8 or NVFP4
- Strong multilingual support and image input
- Available on all major clouds
Watch out for
- No step-by-step reasoning mode
- Low independent score: 9 on the Artificial Analysis Intelligence Index
- Weaker at coding and agents than Mistral Medium 3.5
- Verbose output raises real costs
Specs at a glance
| Developer | Mistral AI |
|---|---|
| Release date | 2 December 2025 (API version 25.12) |
| Architecture | Granular mixture of experts: 673B language model plus 2.5B vision encoder |
| Parameters | 675B total, 41B active |
| Context window | 256K tokens |
| Input / output | Text and images in, text out |
| Reasoning | No built-in step-by-step reasoning (instruct model) |
| Languages | Multilingual; Mistral cites 40+ languages |
| Training | Trained from scratch on about 3,000 NVIDIA H200 GPUs |
| Licence | Apache 2.0 |
| Hardware | FP8: one node of 8x H200. NVFP4: one node of H100s or A100s |
| Where to get it | Mistral AI Studio, Hugging Face, Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks, Together AI, Modal |
| Speed | About 77 output tokens per second (Artificial Analysis) |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| LMArena (at launch) | #2 open non-reasoning model; #6 open model overall | Mistral AI | Vendor-reported, December 2025 |
| GPQA Diamond | 67.2% | Hugging Face model card evaluation | |
| Artificial Analysis Intelligence Index | 9 | Artificial Analysis | Independent, September 2026 |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| Mistral API | $0.50 in / $1.50 out per 1M tokens | Batch 50% off; cached input discounts available |
| Self-host | Free | Apache 2.0; one 8-GPU server |
| Cloud marketplaces | Varies | Amazon Bedrock, Azure Foundry, IBM watsonx and others set their own prices |
Large 3 vs Medium 3.5, which Mistral model?
| Mistral Large 3 | Mistral Medium 3.5 | |
|---|---|---|
| Released | Dec 2025 | Apr 2026 |
| Design | 675B MoE, 41B active | 128B dense |
| Reasoning mode | No | Yes, adjustable |
| Context | 256K | 256K |
| API price (in / out) | $0.50 / $1.50 | $1.50 / $7.50 |
| AA Intelligence Index | 9 | 14 |
| SWE-bench Verified | Not published | 77.6% |
| Licence | Apache 2.0 | Modified MIT (commercial licence above $20M monthly revenue) |
The newer, smaller Medium 3.5 is the better model. Large 3 wins on price (five times cheaper on output) and on licence freedom.
Running it yourself
Mistral publishes the weights on Hugging Face in several formats:
- FP8 (the main post-trained version): one node of 8x NVIDIA H200 GPUs.
- NVFP4 (4-bit): one node of H100s or A100s.
- BF16: full precision for research and fine-tuning.
Mistral recommends vLLM with the model split across 8 GPUs. Tools such as Unsloth support fine-tuning. If you do not have that hardware, the API or a cloud marketplace will be cheaper.
Where it fits in 2026
When it launched, Large 3 was among the strongest open models without a reasoning mode. Since then, open models with reasoning, such as GLM-5.3, MiniMax M3 and Inkling, have moved well ahead on benchmarks. Large 3's remaining edge is its clean Apache 2.0 licence, European origin and simple deployment. See our best open-source LLMs ranking for the full field.
- Use Large 3 for translation, summarising and multilingual support bots, where its lack of a reasoning mode matters less.
- Ask for short answers in the system prompt; Artificial Analysis found it verbose, and output tokens cost three times input.
- Self-host the NVFP4 version if you only have H100 or A100 GPUs; it fits on a single node.
- For non-urgent bulk jobs, use Mistral's batch API to halve the price.
Jargon explained
- Mixture of experts (MoE)
- A design where only a small part of the model switches on for each word, so a very large model runs more cheaply.
- Non-reasoning model
- A model that answers straight away instead of first writing out hidden step-by-step thinking. It is faster but weaker on hard problems.
- FP8 / NVFP4
- Compressed number formats (8-bit and 4-bit) that shrink a model so it needs fewer or cheaper GPUs.
- GPQA Diamond
- A set of very hard science questions written by PhD experts, used to test reasoning.
Alternatives to consider
Frequently asked questions
When was Mistral Large 3 released?
On 2 December 2025, as part of the Mistral 3 family alongside the smaller Ministral 3 models.
How much does Mistral Large 3 cost?
$0.50 per million input tokens and $1.50 per million output tokens on Mistral's API, as of 23 September 2026. The weights are free under Apache 2.0.
What hardware do I need to run Mistral Large 3?
One server with 8 NVIDIA H200 GPUs for the FP8 version, or one node of H100s or A100s for the 4-bit NVFP4 version.
Is Mistral Large 3 a reasoning model?
No. It answers directly without a step-by-step thinking phase. Mistral's reasoning-capable general model is Medium 3.5.
Is Mistral Large 3 better than Medium 3.5?
No. Despite its name and size, Medium 3.5 scores higher on independent tests (14 vs 9 on the Artificial Analysis Intelligence Index). Large 3 is cheaper and has a more permissive licence.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing Mistral 3 (Mistral AI)
- Mistral-Large-3-675B-Instruct-2512 model card (Hugging Face / Mistral AI)
- Mistral Large 3: model analysis (Artificial Analysis)
- Models overview (Mistral AI)
- Pricing (Mistral AI)