证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
self_hosted FP8 instruct weights; NVFP4 on a single node of H100s or A100s; BF16 variant available
高置信度，已绑定来源与证据。
条件：Re-verified on 2026-10-20; the card documents self-hosted deployment of the FP8 instruct weights plus NVFP4 and BF16 variants. Serving frameworks are not specified on the card.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
lifecycle_status
公开事实
active
高置信度，已绑定来源与证据。
条件：Set on 2026-10-20 because the mistralai Hugging Face repository remains public, not gated and not disabled; Mistral does not publish a formal lifecycle table for these weights, so this reflects current availability rather than a vendor-stated status.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
release_date
公开事实
2025-12
高置信度，已绑定来源与证据。
条件：Set on 2026-10-20 from the official repository name suffix 2512; the model card states only the December 2025 release, so day precision is not asserted.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.
条件：Re-verified on 2026-10-20; the official mistralai repository publishes the instruct weights (FP8 plus NVFP4 and BF16 variants) publicly and is not gated.
证据：Mistral Large 3 is a multimodal granular Mixture-of-Experts model with 41B active parameters and 675B total parameters trained from the ground up with 3000 H200s. This instruct post-trained version is published in FP8 and is fine-tuned for chat, agentic and instruction use cases. Key features: a 2.5B vision encoder enabling image analysis in addition to text; agentic capabilities with native function calling and JSON output; Apache 2.0 license; 256k context window. Front matter: license: apache-2.0. Deployment notes cover NVFP4 on a single node of H100s or A100s and a BF16 release. The Hugging Face model API reports the repository as public and not gated.