Codestral Mamba Architecture

7/30/2025

This post announces Codestral 25.08 with specific performance improvements in completion acceptance, code retention, and generation stability. It also introduces the broader Mistral Coding Stack for Enterprise, which integrates Codestral Embed for codebase-scale search, Devstral for agentic workflows, and Mistral Code for IDE integration. Key enterprise-focused features include self-deployment options (cloud, VPC, on-prem), no mandatory telemetry, SSO, audit logging, and usage observability via the Mistral Console. The post highlights how these components work together for a full-stack AI-native software development solution.

6/4/2025

Introduces Mistral Code, an enterprise AI coding assistant that bundles Codestral, Codestral Embed, Devstral, and Mistral Medium models with an in-IDE assistant and enterprise tooling. It enhances the Continue project with enterprise-grade features like RBAC, audit logging, and usage analytics. The offering supports local deployment (on-prem, air-gapped), model fine-tuning/post-training on private data, and multi-step coding tasks. It provides a vertically-integrated solution with unified SLAs for enterprise adoption.

5/21/2025

Introduces Devstral, an agentic LLM specifically trained for software engineering tasks. Highlights its performance on SWE-Bench Verified, outperforming existing open-source and some closed-source models. Details its versatility for local deployment, enterprise use, and integration into coding environments. Mentions its availability under Apache 2.0 and on various platforms, including Mistral AI's API.

7/16/2024

Introduces Codestral Mamba, a new model based on the Mamba architecture, highlighting its linear time inference and ability to model long sequences. Details its training for code and reasoning capabilities, benchmarked up to 256k tokens for in-context retrieval. Provides deployment instructions via mistral-inference SDK and TensorRT-LLM, and mentions availability of raw weights on HuggingFace. Contrasts its Apache 2.0 license with Codestral 22B's commercial/community licenses.

5/29/2024

This post introduces Codestral, a new 22B parameter code generation model. It details its training on over 80 programming languages, its fill-in-the-middle capabilities, and its performance benchmarks across Python, SQL, and other languages, highlighting its superiority in RepoBench due to a 32k context window. It also announces the availability of a dedicated `codestral.mistral.ai` API endpoint, its integration with popular developer tools and frameworks, and licensing details.