Mistral Large 3 suits teams that want a self-hostable, openly licensed alternative to closed frontier models for coding and long-document work, scoring about 92% on HumanEval. It's the wrong pick for graduate-level science reasoning, where newer open-weight reasoning models pull well ahead, and for teams lacking the multi-GPU infrastructure needed to self-host it at scale.
Mistral Large 3 ranks #2 among open-weight non-reasoning models on LMArena with an Elo near 1418, marking Mistral AI's return to Mixture-of-Experts architecture. It pairs multimodal image understanding with native function calling, positioning it as a self-hostable alternative to closed frontier coding models.
Provider: Mistral AI · Family: Mistral Large
More about Mistral AI on HokAI
Context window: 256,000 tokens · Max output: 256,000
Input modalities: text, image · Output: text
About Mistral Large 3
Mistral Large 3 (API name mistral-large-2512) is Mistral AI's flagship general-purpose model, released on December 2, 2025. It marks a return to Mixture-of-Experts architecture after the original Mixtral series, using a granular MoE design with 41 billion active parameters out of 675 billion total, trained from scratch on 3,000 NVIDIA H200 GPUs. It succeeds Mistral Large 2 and launched alongside the wider Mistral 3 family, released under an open license with downloadable weights, unlike Mistral Medium 3, which remains proprietary and API-only. On benchmarks, Large 3 scores roughly 85.5% on an 8-language MMLU evaluation and reaches about 92% pass@1 on HumanEval, close to GPT-4o's coding performance. Its LMArena Elo of approximately 1418 ranks it #2 among open-weight non-reasoning models on the leaderboard. Its clearest weak spot is graduate-level science reasoning: GPQA Diamond comes in at 43.9%, well behind current open-weight reasoning leaders. Mistral has not published SWE-bench Verified, AIME 2025, or ARC-AGI 2 scores for this model. Large 3 ships with a context window more than double Mistral Medium 3's, engineered for long documents, codebases, and multi-turn agent sessions without truncation. It is multimodal, accepting text and up to 8 images per request, supports document OCR through chat completions, and natively handles 40+ languages, backed by function calling and Mistral's built-in tools framework for agentic workflows. As open-weight, it is downloadable on Hugging Face in base and instruct forms, and also available through Amazon Bedrock, Microsoft Azure AI Foundry, IBM watsonx, Google Cloud Vertex AI, Fireworks, Together AI, and OpenRouter. Self-hosting is demanding: because the MoE router must read every expert's weights to select which subset activates per token, the full model must fit in VRAM, requiring roughly 355GB at 4-bit quantization or double that at FP16, typically spread across multiple GPU nodes with vLLM and expert parallelism. Mistral AI is based in Paris and operates under EU data protection law; governance documentation is published on its AI Governance Hub, though no model-specific system card with refusal-rate figures has been released. A dedicated reasoning variant was announced alongside the December 2025 launch but had not shipped as of Mistral's most recent product updates.
Pricing
Mistral Large 3 is priced at $0.50/$1.50 per million tokens (input/output) on Mistral's own La Plateforme API, confirmed in Mistral's official documentation, with matching rates on Amazon Bedrock and Azure AI Foundry. Self-hosting the open weights costs nothing beyond your own GPU infrastructure.
Key Features
- Granular Mixture-of-Experts Routing: Only a fraction of the total parameters activate per token, keeping latency and cost close to a much smaller dense model while retaining frontier-scale capacity.
- 256,000-Token Context Window: Handles long documents, full codebases, or extended multi-turn agent sessions in a single request without chunking or truncation.
- Multimodal Input, Up to 8 Images: Accepts text alongside as many as 8 images per request for cross-modal analysis, plus document OCR through the chat completions API, across 40+ languages.
- Native Function Calling and Tool Use: Supports Mistral's built-in tools framework for multi-step agentic workflows, with structured JSON output built in.
Pros
- Ships under a permissive open license with no commercial-use restrictions, unusual for a model at this scale.
- Roughly matches GPT-4o-level coding ability on HumanEval while costing a fraction of most closed frontier models per token.
- A context window built for whole-codebase or long-document analysis without splitting content into chunks.
Cons
- GPQA Diamond of 43.9% trails DeepSeek V4's 90.1% by a wide margin on graduate-level science and math reasoning.
- Self-hosting needs about 355GB of VRAM at 4-bit quantization to hold all parameters in memory, even though only a fraction activate per token.
- No configurable reasoning-effort mode shipped at launch; Mistral's other 2026 models added one but Large 3 has not yet.
Benchmarks
- mmlu: 85.5
- mmlu pro: 81
- humaneval: 92
- lmarena elo: 1418
- gpqa diamond: 43.9
- artificial analysis speed tokens per sec: 52
Frequently Asked Questions
How much does Mistral Large 3 cost in 2026?
Mistral Large 3's API pricing is $0.50 per million input tokens and $1.50 per million output tokens, the same across Mistral's own platform, Amazon Bedrock, and Azure AI Foundry. A long-document workload processing 1 million input and 100,000 output tokens a day runs about $0.65; a heavier multimodal agent loop costs closer to $2.40 a day. Because the weights are open, self-hosting is free aside from compute.
Is Mistral Large 3 open source?
Yes. Mistral Large 3 ships under the Apache 2.0 license with no restrictions on commercial use, modification, or redistribution. Both base and instruction-tuned checkpoints are downloadable from Hugging Face, including quantized variants for smaller deployments. Because there is no hosted free tier, the only no-cost path is self-hosting on your own GPU infrastructure.
What are the best alternatives to Mistral Large 3?
For graduate-level reasoning work, DeepSeek V4 outperforms Large 3 on science and math benchmarks at a comparable price. For a cheaper proprietary option with a shorter context window, Mistral Medium 3 costs less per output token. For teams that want the strongest open-weight reasoning score available today, Kimi K3 leads, at a steeper price.
How does Mistral Large 3 compare to DeepSeek V4 in 2026?
DeepSeek V4 costs $0.435 per million input tokens and $0.87 per million output tokens, slightly less than Large 3, and ships under the MIT license. It also posts far stronger graduate-level science and math reasoning scores. Large 3 answers with a more established multi-cloud presence across Bedrock, Azure, and Vertex AI, and a smaller, better-documented self-hosting footprint.
How do you get started with Mistral Large 3?
Create an API key on Mistral's La Plateforme and send a chat completions request using the mistral-large-2512 model name; quickstarts are at docs.mistral.ai. Teams already on AWS, Azure, or Google Cloud can instead call it through Bedrock, Azure AI Foundry, or Vertex AI with existing cloud credentials. To self-host, download the instruct checkpoint from Hugging Face and deploy with vLLM using expert parallelism across a multi-GPU node.
Top Alternatives
- DeepSeek V4: Pick DeepSeek V4 if graduate-level science and math reasoning matters most: it posts a 90.1% GPQA Diamond score at a comparable per-token price.
- Mistral Medium 3: Pick Mistral Medium 3 if a shorter context window is enough and you want a proprietary API; pick Large 3 for double the context and open weights you can self-host.
- Kimi K3: Pick Kimi K3 for the highest open-weight GPQA Diamond score at 93.5%; pick Large 3 for a lower per-token API price and simpler deployment.