Mistral has released open weights for Mistral Large 4, its 1 trillion-parameter AI model for tasks including automated coding. A quantised version reportedly runs on 4 Nvidia B200 or B300 GPUs, giving builders a hardware configuration to assess for running it themselves.
Mistral Large 4 is now available on Hugging Face and through Mistral’s API. It accepts images and text as input and generates code or text, using 52 billion active parameters per task.
The release follows a preview 3 weeks earlier, with further reinforcement learning guided by customer feedback. Mistral reports that results on Terminal-Bench 4.0, which tests coding agents in a terminal, rose from 28% to 41%, compared with GLM-5.3 at 40%.
For builders, token use and control over reasoning are worth testing alongside coding accuracy. Results reported by Mistral on the Coding Agent Index from Artificial Analysis put the model alongside GLM-5.3, with average token consumption per task 47% lower; Mistral has also added low and medium reasoning modes, but API pricing has not yet been published.
Released weights give companies an alternative to relying solely on a hosted model whose access they do not control.
Lower reported token use makes Mistral Large 4 worth evaluating for long-running coding agents, although actual cost savings remain unproven.
The quantised version gives infrastructure teams a specific hardware configuration to assess, rather than a vague promise of self-hosting.
Written automatically by the Made in 10 Newsroom from the source above.
Deepgram is acquiring Daily, the company behind the open-source voice AI framework Pipecat. The companies say Pipecat will remain open-source, preserving developers' ability to connect different models.
Google Chrome is rolling out AI autofill that uses Gemini to suggest information from Gmail when users complete online forms. The feature starts in the US and requires Personal Intelligence to be enabled.
GitHub launched Copilot Fusion to choose which AI model, or combination of models, handles each stage of a coding task. GitHub Enterprise customers can also access every Copilot feature without a separate subscription, paying for the AI credits they use.