May 29, 01:00 AM

Mistral AI Launches Codestral Embed, Open-Source Model Beating OpenAI and Cohere, Runs on RTX 4090 or Mac

Mistral AI has launched Codestral Embed, its first embedding model specifically designed for code. The model outperforms leading competitors including OpenAI's Text Embedding 3 Large, Voyage Code 3, and Cohere Embed v4.0 in real-world retrieval tasks. Codestral Embed is fully open-source under the Apache 2.0 license, allowing commercial use. It surpasses other models such as Gemma 3 27B and DeepSeek V3 on the SWE-Bench Verified benchmark. The model is optimized for agent use and can run locally on a single RTX 4090 GPU or a Mac with 32GB of RAM. This release marks a notable development in AI code embedding technology. Separately, the Instella model suite, a fully open 3 billion parameter model trained on AMD Instinct GPUs, was also announced, emphasizing transparency and reproducibility in AI development.