Mistral releases Le Chonk weights, with a license caveat for model services

Mistral’s October 29 announcement releases the weights of an improved Mistral Large 4, nicknamed Le Chonk. The company says additional reinforcement learning improved coding and defensive cybersecurity after its earlier preview. Those gains are company claims; this desk has not independently tested them. [1]

The model card describes text-and-image input, text output, a one-million-token context window and low, medium and high reasoning settings. It lists roughly one trillion total parameters, with about 52 billion active including embeddings and output layers. Its serving examples use eight GPUs; that is a documented configuration, not our minimum-hardware test. [2]

Mistral lists Studio and its API as hosted routes, alongside downloadable weights for deployment with vLLM or SGLang. Its documentation currently prices a million tokens at $0.70 for input, $0.07 for cached input and $2.10 for output. Downloading weights and paying for hosted inference are separate propositions. [1][4]

The Mistral Ecosystem Open-Weight License is modified MIT. It requires preserving its notices. If a licensee or affiliate operates a Model-as-a-Service business, a separate Mistral commercial license is required before commercial use. The definition excludes certain embedded end-user features and simple request relaying; clause 3 exempts internal use and access through Mistral’s official products from that requirement. [3]

Spicy's takeThe practical change is a choice about where the model runs and how it can be adapted. A buyer can evaluate hosted access against operating the downloadable model under its terms. That is useful flexibility; the nickname does not pay the infrastructure bill. Compare deployment effort, workload fit and licensing before treating a benchmark claim as a purchasing decision. [1][2][3]

Spicy's takeMistral says a full technical report is still forthcoming. This release establishes availability and documented options, rather than proving superior results on a reader’s own tasks. Our review covered the announcement, model card, license and current documentation; it did not include running the model. [1][2]