Mistral Devstral 2: The DeepSeek Killer we wanted?

Here’s a summary of the provided YouTube transcript:

Mistl has released Mistral Vibe, a new coding model and CLI, alongside two new models, Devstral 2 and Devstral small, aiming to compete with the openweight ecosystem. The model is based on the Llama architecture and is designed to be cost-effective, offering a 7x efficiency boost compared to Cloud Sonnet. Devstral 2 is a 123 billion parameter model, while Devstral small is a 24 billion parameter model, with the smaller model being suitable for consumer hardware. The model is released under a modified MIT license, and the larger model is priced at $40/million input tokens and $2/million output tokens. The Mistral CLI is a VIP coding tool integrated with ZAI’s agent communication protocol, enabling IDE integration and offering a tool called Success Rate, which is comparable to closed-source models. The Mistral Vibe CLI is available through the ZED IDE, allowing users to export the model and run it in the command line. The initial release of Devstral 2 and Devstral small has generated mixed feedback, with the larger model being closely followed by Deepseek, but the toggle button is not working. The video explores the licensing model, highlighting the 123B parameter model’s MIT license and the 24B parameter model’s Apache 2.0 license. The video also discusses the benchmark results, demonstrating that Devstral 2 can deliver up to seven times more cost-efficient coding compared to Cloud Sonnet, with the larger model being comparable to Deepseek’s version 3.2. The video also touches on the development of the Mistral Vibe CLI, which is a valuable tool for integrating with IDEs and offers a set of features like status, steps, and support for MCP. The video concludes by encouraging viewers to test the model and provides a discussion of the community’s perspective on the release.