French AI startup Mistral has unveiled a new open-weight AI model, Mistral Large 4, with 1 trillion parameters. CNBC reported on Oct. 6 that Mistral said Large 4 is among the world’s best open-weight models and is particularly strong at cyber, manufacturing, finance and multimodal tasks. It acknowledged it still falls short of frontier models in some areas such as coding.

Mistral plans to provide ML4 first as a preview to developers, “cybersecurity leaders” and national authorities, and to release it more broadly in late October.

ML4 was trained for two months using 4,000 Nvidia Grace Blackwell GPUs installed at the company’s data centres in Europe.

Mistral said that if it discloses key parameters, ML4 would rank among the world’s top open-weight models based on overall benchmark performance. It also claimed it is the most powerful by a wide margin among open-weight models developed outside China.

Guillaume Lample (기욤 랑플), Mistral’s co-founder and chief scientist, said the model’s performance will improve further as the company expands training capacity after raising Series D funding. He added that it would help companies and governments respond in cyber defence capabilities.

Competition in open models has emerged as a key issue in the U.S.-China technology rivalry. Open models allow users to modify them directly or run them on their own infrastructure, distinguishing them from closed models offered by OpenAI and Anthropic.

Earlier this year, Hugging Face drew attention for using China-based Z.ai’s GLM 5.2 while responding to an unusual OpenAI agent breach. At the time, Hugging Face tried to use a closed model from a U.S. company first, but safety guardrails misidentified breach response as an attack and blocked it.

Mistral raised 3 billion euros in a Series D round in September and was valued at 21 billion euros. U.S. startup Reflection AI, backed by Nvidia, also unveiled its first open model on Oct. 5 and claimed performance comparable to GLM 5.2.