Mistral

Processes both text and image inputs simultaneously within a 256k token context window, enabling analysis of documents with embedded visuals, screenshots with surrounding text, and multi-page content. Mistral Large 3 uses a unified transformer architecture to fuse text and vision embeddings, allowing cross-modal reasoning where image content informs text generation and vice versa. The extended context window (256k tokens ≈ 200 pages) enables processing of entire documents without chunking.

Magistral model exposes its internal reasoning process through explicit reasoning tokens that show step-by-step problem decomposition before generating final answers. This architecture allocates a portion of the token budget to internal reasoning (similar to OpenAI's o1 approach) rather than direct output generation, enabling verification of reasoning quality and debugging of incorrect conclusions. Users can inspect the reasoning trace to understand how the model arrived at its answer.

Mistral Studio is a web-based IDE for building AI agents and applications without writing code. Users define agent behavior through a visual interface, connect tools/APIs, and deploy agents directly. The platform abstracts away prompt engineering and API integration complexity, enabling non-technical users to build functional AI applications. Agents built in Studio can be deployed as APIs or embedded in applications.

Mistral Vibe is a VS Code and JetBrains IDE plugin providing real-time code completion suggestions powered by Codestral. The plugin integrates with the editor's autocomplete system, showing suggestions as the user types. Uses pay-as-you-go pricing (charged per completion request) rather than per-seat subscriptions, reducing cost for teams with variable usage. Supports multiple programming languages and includes context awareness for project-specific patterns.

Le Chat is Mistral's web-based chat interface accessible via browser, offering free and paid tiers. Free tier provides limited access to Mistral models with usage caps. Pro tier ($14.99/month) includes higher usage limits and priority access. Team tier ($24.99/month per user) adds collaboration features. Enterprise tier offers custom pricing and dedicated support. Web interface integrates web search, file uploads, and conversation history without requiring API integration.

Mistral Small 3 achieves 81% accuracy on the MMLU (Massive Multitask Language Understanding) benchmark, a standard evaluation of general knowledge across 57 subjects. This benchmark result is publicly documented and verifiable, providing a concrete performance metric for model quality. MMLU score enables comparison with other models on a standardized scale (GPT-3.5 ≈ 86%, Claude 3 Haiku ≈ 75%, Llama 2 ≈ 45%).

Mistral Small 3 achieves 150 tokens per second inference speed on standard hardware (hardware specification not documented). This throughput metric indicates latency for real-time applications: 150 tokens/sec ≈ 6.7ms per token, enabling sub-second responses for typical queries (100-200 tokens). Speed is likely achieved through optimized inference kernels and efficient model architecture (grouped query attention, etc.).

Codestral 25.01 is a code-specialized model trained with emphasis on code generation, completion, and repair across multiple programming languages. The model uses code-specific tokenization and training objectives optimized for syntax correctness and idiomatic patterns. Integrated into Mistral Vibe (CLI and IDE plugin) for in-editor code suggestions with pay-as-you-go pricing, enabling real-time code completion without subscription overhead.

Mistral Large 3 and Ministral family models support multilingual input and output across 40+ languages with unified tokenization and training. The models use a shared vocabulary and transformer architecture trained on multilingual corpora, enabling code-switching (mixing languages in a single prompt) and translation-adjacent tasks without explicit translation models. No separate language selection required; language is inferred from input.

Document AI model is specialized for extracting structured data from documents including text, tables, and handwritten content. The model uses document-specific training objectives and likely incorporates layout understanding (detecting columns, headers, footers) and optical character recognition (OCR) capabilities. Enables extraction of tabular data into structured formats and recognition of handwritten annotations without separate OCR pipelines.

Ministral family (3B, 8B, 14B parameter variants) is engineered for edge deployment on resource-constrained devices including mobile phones, IoT devices, and embedded systems. Models use parameter-efficient architectures (likely including techniques like grouped query attention, knowledge distillation, or pruning) to maintain capability while reducing memory footprint and inference latency. Enables on-device inference without cloud connectivity, reducing latency to <100ms and eliminating API costs.

Le Chat (Mistral's web interface) integrates web search capability, enabling the model to retrieve and cite current information from the internet before generating responses. The system likely uses a search API (Google, Bing, or proprietary) to fetch relevant documents, embeds them in the context window, and generates answers with source attribution. Enables answering questions about recent events, current prices, and breaking news that are outside the model's training data cutoff.

Mistral Large 3 includes agentic capabilities enabling the model to decompose complex tasks into subtasks, call external tools (APIs, functions), and iterate based on results. The model uses chain-of-thought reasoning to plan tool sequences and can handle tool failures by retrying or switching strategies. Enables building autonomous agents that can accomplish goals requiring multiple API calls and decision-making without explicit orchestration code.

Mistral offers API access with per-token billing model (input tokens and output tokens charged separately) rather than subscription-based pricing. Users pay only for tokens consumed, enabling cost-effective usage for variable workloads. Pricing structure is transparent and documented in the API dashboard, with usage tracking and spending alerts available. No minimum commitment or monthly fees required.

Mistral Small 3 is distributed as an open-weight model under Apache 2.0 license, enabling free download, modification, and commercial use without licensing fees. The model weights are available in standard formats (safetensors, GGUF) for self-hosting on any infrastructure. Apache 2.0 license provides legal clarity for commercial applications and derivative works, with minimal restrictions (attribution required, no liability).