Mistral Releases Large 4 Open-Weight AI Model With 1T Parameters

Mistral has introduced its latest artificial intelligence model, known as Mistral Large 4 or Le Chonk. The system is a multimodal, multilingual, Mixture-of-Experts model that accepts both text and images as inputs and produces text outputs. It features a total parameter count of roughly one trillion, with 49 billion active parameters, and supports a context window of one million tokens. The model is distributed under the Mistral Ecosystem Open-Weight License, which is a derivative of the MIT license that places specific limits on model-as-a-service offerings.

Architecture and Capabilities

The system combines a large pool of experts with sparse attention and a vision encoder. Its text decoder contains 72 layers, a hidden dimension of 6,144, and 64 attention heads. The attention blocks integrate Multi-head Latent Attention with a sparse attention indexer. Out of the 72 layers, the first uses a dense feed-forward network, while the remaining 71 layers use Mixture-of-Experts blocks. A router selects four out of 128 experts for each token, alongside one shared expert, resulting in five active experts per token.

Mistral Large 4 performs particularly well in cybersecurity and coding tasks. It scored 82 percent on a cyber index test that requires models to reproduce and patch real vulnerabilities in open-source software. According to Mistral, several closed frontier models refused to attempt the test and scored near zero. The model also includes a 1.6-billion parameter vision encoder and supports various deployment frameworks, including day-zero support for transformers, vLLM, and llama.cpp.

What It Means for Builders and Investors

For builders and developers, open-weight models like Mistral Large 4 provide greater control over usage policies and local deployment compared to closed APIs. This freedom is especially valuable for defensive security work, where handling malware samples and vulnerability research can trigger safety filters on restricted cloud platforms. Teams can build custom security systems around their own codebases and logs without risking sudden loss of API access during critical incidents.

Investors and researchers gain an open window to evaluate actual model behavior, refusal rates, and safety guardrails directly. The availability of quantized versions and efficient execution frameworks allows organizations to run advanced multimodal capabilities locally or on specialized hardware, lowering barriers to downstream fine-tuning and specialized agentic workflows.