Mistral has released open weights for Mistral Large 4, its trillion-parameter model. In a post on October 29, Hugging Face said the model is available in two precisions, FP8 and NVFP4, and that Transformers, vLLM and llama.cpp could run it on day one. Unsloth shipped GGUF quantized builds at the same time. Mistral had promised an open release by the end of the month when it launched the research preview on October 6, so this lands on schedule.
Architecture and license
The repository is named Mistral-Large-4-1T-A52B. According to Hugging Face, the model has 72 layers and a mixture-of-experts (MoE) design with 128 experts, where each token is routed to 4 experts plus 1 shared expert. The attention block uses multi-head latent attention (MLA) with sparse indexing, so each query looks at no more than 2,048 key positions. On the vision side, a 1.6B Pixtral encoder handles image input, and the model takes text plus images in and produces text out.
There are currently two figures for active parameters: the blog post says about 49 billion, while the repository name says 52 billion. Mistral's follow-up clarification should settle it. The context window is rated at 1 million tokens, but real deployment configurations are lower: the vLLM and Inference Endpoints examples are set to 131K, and the 2-bit llama.cpp build runs at 800K.
The license is the Mistral Ecosystem Open-Weight License, a rewrite of MIT that restricts packaging the model directly as a model-as-a-service offering. Personal use, fine-tuning and private deployment are unaffected, but cloud providers planning to resell it as an API will want to read the terms first.
Scores cluster around security and coding
Among the results Hugging Face lists, security stands out most: 93% on the 40 security competition tasks in Cybench, 82% on vulnerability reproduction and patching, 93.3% on Lakera B3 attack resistance, and a top-five place worldwide on the Artificial Analysis cybersecurity index. For coding agents, it scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. Its overall intelligence index ranks fifth among open-weight models.
"Mistral Large 4 is particularly strong for cybersecurity, coding agents, and multimodality."
These figures come from Mistral and its partners. How much of this the community can reproduce with the weights in hand will take time to show.
From Large 3 to open weights
Put the Mistral line side by side and the change is considerable. According to Artificial Analysis data published at the time of the preview, Large 3 scored just 9 on its intelligence index, while the Large 4 preview reached 38. During the preview the model was available only through Mistral's own API, prioritized for developers, security leads and government agencies, and Mistral acknowledged then that it still trailed the most advanced closed models in areas such as coding. The 28.3% on Terminal-Bench 4 published this time sits well below the security scores above 90%, which broadly matches that self-assessment.
With the weights out, deployment options have multiplied. Hugging Face Inference Endpoints uses 8 H200 GPUs, the Dell Enterprise Hub example uses 4 B300s, and the TRL fine-tuning example needs one 8-GPU B300 node. The 2-bit llama.cpp quantization fits on a single Apple Silicon Ultra machine with 512GB of unified memory, while the other quantizations need distributed multi-GPU runs, which the official team tested on 4 RTX 6000 Pro cards.
When we covered the preview, we estimated that storing the weights alone at 8-bit precision would take about 1TB of GPU memory. NVFP4 and 2-bit builds have since lowered the bar, but for most teams an 8-GPU Blackwell server is still the starting configuration. Official pricing on the Mistral API was not mentioned in this announcement.
Sources: Hugging Face official blog, vLLM blog, CocoLoop, Artificial Analysis; parameter structure, context length and license name checked against the Hugging Face model card.