Pixtral Large

Pixtral Large, released on 18 November 2024, was the moment Mistral put a frontier-scale multimodal model into the open-weights world. It was built as a 124-billion-parameter system: the Mistral Large 2 text decoder with a 1-billion-parameter vision encoder bolted on top, so the model read images without giving up the language ability of the flagship it was based on. Its strength was documents rather than photographs. Mistral reported 69.4% on MathVista — reasoning over charts and diagrams — and claimed wins over GPT-4o and Gemini 1.5 Pro on DocVQA and ChartQA, the benchmarks that matter for invoices, reports and scientific figures. On MM-MT-Bench, a multimodal instruction-following test, it was reported ahead of Claude 3.5 Sonnet, Gemini 1.5 Pro and GPT-4o. The 128,000-token context window was specified as holding at least thirty high-resolution images at once, which made whole-document work practical rather than page-by-page. The licensing was the split Mistral used throughout this period: weights on Hugging Face under the Mistral Research License for research and education, and a separate Mistral Commercial License for anything commercial — open to inspect and study, not free to sell on. Mistral deprecated the model on 27 February 2026 and switched it off on 31 May 2026, pointing users to Mistral Medium 3.5. The capability did not disappear so much as get absorbed: vision stopped being a separate model at Mistral and became a property of the mainline ones.

▸Videos

No videos yet.

▸Reviews