French artificial intelligence (AI) startup Mistral AI on Thursday released the open weights for Mistral Large 4, its latest flagship frontier model tailored to enterprise tasks and defensive cybersecurity.
Nicknamed Le Chonk (internet slang for “big and cute”), the multimodal model arrives following three weeks of public preview testing with corporate clients and government agencies. Unlike proprietary systems such as Google’s Gemini that operate solely behind cloud application interfaces, Mistral Large 4’s underlying weights can be downloaded directly from platforms like Hugging Face, allowing organizations to self-host and modify the system on their own private servers.
Company executives framed the rollout as a bid for technological sovereignty.
“I’m not comfortable living in a world in which open weights do not exist,” said Pierre Stock, Mistral’s vice president of science, noting that downloadable weights insulate enterprises from vendor lock-in and sudden service cutoffs.
Marjorie Janiewicz, Mistral’s chief revenue officer, echoed that demand, citing enterprise priorities around customization, internal data governance, and overall business continuity.
Architecturally, the model is a trillion-parameter, sparsely activated system featuring 1.05 trillion total parameters, 52 billion active parameters per forward pass, and a 1-million-token context window. Trained on roughly 3,800 NVIDIA Grace Blackwell GPUs across more than 160 languages, the system processes both text and imagery.
Mistral emphasized its agentic capabilities for complex multi-hour workflows, its satellite and aerial image analysis, and its robust grounding techniques designed to reduce hallucinations. In software evaluation environments such as SWE-bench and Terminal Bench, the system demonstrated strong aptitude for code auditing, pull requests, and threat detection.
Independent evaluations from benchmarking firm Artificial Analysis place Mistral Large 4 at 26 points on its Intelligence Index v5 under high reasoning mode, edging out Western peers like Meta Platforms Inc.’s Muse Spark Open. The achievement makes it the highest-performing open-weight model developed outside China, though it still trails Chinese open-weight leaders DeepSeek V4.1 Flash and GLM-5.3.
The update also brings sharp economic improvements. Artificial Analysis reported a 63% decline in benchmark operating costs relative to the preview version, dropping from $5.54 to $2.06 per indexed task. Mistral halved its direct API rates alongside the launch to $0.70 per million input tokens and $2.10 per million output tokens, though running the massive architecture locally will still demand significant enterprise compute investments.
The release is governed by the Mistral Ecosystem Open-Weight License, a custom adaptation of the MIT open-source standard. While granting broad rights to run, adapt, and distribute the weights without arbitrary revenue limits, the license enforces a targeted restriction: entities operating model-as-a-service platforms that sell direct access or post-training capabilities to third parties must negotiate a separate commercial agreement. Routine internal deployments and embedded enterprise applications remain exempt, safeguarding general business adoption while protecting Mistral’s commercial ecosystem.
