Mistral releases ML4, a 1-trillion parameter open-weight model
Mistral has released the weights for ML4, its largest model to date, featuring 1 trillion total parameters and trained entirely on its own French data centers. The model is positioned as a competitive alternative to closed-source US models, emphasizing sovereignty and efficiency.
Mistral has released the weights for ML4, its largest model to date, featuring 1 trillion total parameters with 52 billion active parameters per token. The model is designed as a Mixture-of-Experts architecture to balance training complexity with inference efficiency, allowing it to run on four NVIDIA GB300 GPUs when quantized to FP4. ML4 is currently the best open-source model for coding and cybersecurity, solving 18 of 19 tasks in the CTF Speedrun benchmark and blocking 9 out of 10 prompt injections in internal tests. The open-weight version uses 50% fewer tokens than the preview version released three weeks prior while maintaining the same intelligence level, making it the most efficient model on the Artificial Analysis Coding Index. Mistral states that ML4 outperforms Opus 4.8 and matches the level of GPT-6 Luna, positioning it as a competitive alternative to closed-source US models.
Mistral trained ML4 entirely on its own data centers in France, utilizing only 10 megawatts of power, which the company describes as significantly less than competitors. This infrastructure was funded by a Series B round, while a recent Series D round of 3 billion euros in September will fund future compute expansion. The company now values at over 21 billion euros, making it Europe's most valuable AI company. Mistral is expanding its platform to offer bare-metal GPU rentals and inference for third-party models, including Chinese models like GLM 5.3 from ZAI, to drive demand and fund further research. The company emphasizes that owning infrastructure allows for faster debugging and deployment, citing an anecdote where pre-training was stabilized in one night due to co-located engineering teams, a process that would take weeks with cloud providers. Mistral contends that building sovereign infrastructure in Europe is a strategic necessity to avoid geopolitical dependency, comparing access to AI tokens to access to electricity. "Ein Vorteil offener Gewichte ist, dass niemand einem den Zugang abschneiden kann," the company stated.