In brief: Mistral has released Large 4 as an API preview. The multimodal mixture-of-experts model contains 1.05 trillion parameters while activating 49 billion per processing step. An independent cyber test confirms a leading result, but the promised open weights are not available yet.
What is available today
Mistral Large 4 supports a one-million-token context window and includes a 1.6-billion-parameter vision encoder. Mistral says it trained the model in Europe on 3,800 NVIDIA Grace Blackwell GPUs and covered more than 160 languages. For now, users can access the public API preview.
Mistral says the weights will arrive by month-end; Reuters specifies 27 October. No licence has been published. Large 4 is therefore more accurately described as a planned open-weight release, not a model that can already be downloaded or operated under known terms.
The cyber comparison is narrower than the headline
The preview scores 81.7% on CyberGym-E2E-AA, ahead of MiMo-V2.6-Pro at 78.6% and GPT-6 Luna at 77.9%. Artificial Analysis runs the evaluation independently: an agent must find, reproduce and patch real memory-safety flaws across 131 open-source projects without breaking existing tests.
This supports Mistral's claim of strong defensive cyber capability, but not a general victory over Chinese models. In the three-test Cyber Index, Grok 4.7 and MiMo-V2.6-Pro lead at 56, followed by GPT-6 Luna at 53; Mistral only says Large 4 ranks in the top five. Reuters correctly notes that CEO Arthur Mensch named neither models nor benchmarks for his broader China comparison.
49 billion active does not mean 49 billion stored
The MoE design reduces compute per token because only some experts are active. A complete self-hosted deployment must still store or distribute all 1.05 trillion parameters. At four bits, the raw weights alone work out to roughly 525 gigabytes, before runtime buffers, context cache and other data. Mistral says detailed architecture information will arrive with the weights.
Pandorex Analysis
Large 4 gives Mistral's strategy of European infrastructure, capital and open weights a concrete technical result. The CyberGym lead is stronger evidence than a vendor-only chart. Whether this becomes a practically open frontier model depends on the 27 October release: weights, licence, reproducible configurations and memory requirements must be assessed together.