Model Customization and Agents

3/17/2026

This post introduces Forge, a new system for enterprises to build AI models grounded in proprietary knowledge. It details Forge's capabilities in training models on internal data (documentation, codebases, records) using pre-training, post-training, and reinforcement learning. It highlights support for dense and MoE architectures, multimodal inputs, and agent-first design. The system emphasizes continuous improvement through RL and evaluation, enabling custom models for enterprise agents to improve reliability, tool use, and decision-making within organizational constraints. Examples of enterprise applications are provided across various sectors.

3/16/2026

Introduces Leanstral, an open-source code agent specifically for Lean 4. Details its sparse architecture (6B active parameters), evaluation methodology using the FLTEval suite, and comparative performance against other coding agents and open-source models. Provides case studies demonstrating its ability to answer Stack Exchange questions about Lean 4 changes and translate/reason about programs from Rocq. Highlights its availability through Mistral Vibe, a free API, and as downloadable weights.

3/11/2026

This post details the technical implementation of an autonomous agent for generating and improving RSpec tests in Rails applications. It describes the agent's architecture, which leverages Mistral's Vibe platform, including the use of `AGENTS.md` for step-by-step execution plans and specialized `SKILLS` files for different Rails file types (models, controllers, etc.). The post also elaborates on the integration of custom tools like RuboCop for linting and SimpleCov for test validation and code coverage, including stripped-down code examples for tool implementation. It further discusses metrics for measuring test quality, including tool-based metrics and the concept of 'LLM-as-a-judge' for qualitative assessment.

8/1/2025

This post details the application of LoRA fine-tuning to Pixtral-12B for satellite imagery classification using the Aerial Image Dataset (AID). It demonstrates how LoRA can efficiently adapt model weights to specific tasks and domains, leading to significant improvements in classification accuracy (from 0.56 to 0.91) and a reduction in hallucinations. The post also highlights the ease of fine-tuning via Mistral's API and LaPlateforme UI, and provides guidance on hyperparameter selection. It showcases the effectiveness of domain-specific adaptation for specialized visual domains like satellite imagery.

7/10/2025

Introduces Devstral Small 1.1 and Devstral Medium, enhancing agentic coding capabilities. Devstral Small 1.1, an upgrade to the previous version, is released under Apache 2.0 and achieves 53.6% on SWE-Bench Verified, showing improved generalization and support for Mistral function calling and XML formats. Devstral Medium, available via API, scores 61.6% on SWE-Bench Verified and offers competitive cost/performance, with options for on-premise deployment and custom fine-tuning.

2/17/2025

Introduces Mistral Saba, a new 24B parameter regional language model focused on the Middle East and South Asia, with strong capabilities in Arabic and South Indian languages. Highlights its performance, cost-efficiency, and local deployability, positioning it as a specialized regional adaptation that can be further fine-tuned.

1/30/2025

Introduces Mistral Small 3, a 24B-parameter model optimized for latency and local deployment. Highlights its performance benchmarks against larger models and proprietary alternatives, and details its suitability for conversational AI, function calling, fine-tuning, and local inference. The post also reiterates Mistral AI's commitment to Apache 2.0 licensing for general-purpose models and mentions upcoming releases with boosted reasoning capabilities.

1/13/2025

Codestral 25.01 introduces a more efficient architecture and an improved tokenizer, leading to a 2x speed improvement in code generation and completion. It achieves SOTA performance for FIM use cases and offers a 256k context length. Benchmarks show significant improvements over previous versions and competitors across various coding tasks and languages.

8/7/2024

This post announces the ability to customize Mistral AI models (including Mistral Large 2 and Codestral) on La Plateforme via base prompts, few-shot prompting, or fine-tuning with user-provided datasets. It also introduces an alpha release of Agents, a feature that wraps models with context and instructions to create custom behaviors and workflows, with future plans for tool and data source integration. The release of mistralai client SDK v1.0 for Python and Typescript is also highlighted.

6/5/2024

Introduced mistral-finetune, a lightweight and efficient codebase for fine-tuning Mistral's open-source models using the LoRA training paradigm. Launched serverless fine-tuning services on la Plateforme, leveraging Mistral's proprietary fine-tuning techniques for fast and cost-effective model adaptation and deployment, also using LoRA adapters. Announced custom training services for fine-tuning models on customer-specific applications with proprietary data, including continuous pretraining.

12/11/2023

This post announces the beta launch of 'La Plateforme', Mistral AI's first platform services, providing API endpoints for text generation and embeddings. It details the models served by each endpoint (mistral-tiny, mistral-small, mistral-medium, mistral-embed), their performance benchmarks (MT-Bench, MTEB), and the underlying alignment techniques used. It also mentions API specifications, client libraries, and the ramp-up process towards general availability, acknowledging NVIDIA's contribution to optimizing sparse mixture of experts models.