Mistral AINew

Updated

France's Mistral models — the Mistral Large 3 flagship is a 675B/41B sparse MoE with a 256K context under Apache 2.0 — plus the Le Chat assistant and a pay-per-token API

  • Popularity
  • Chat
  • Free tier
Visit website
Mistral AI thumbnail
Report incorrect information

Choose an issue below. You do not need to sign in or leave contact details.

At a glance

  • Free tierPartial
  • Open sourcePartialLarge 3 and Small 4 ship as Apache 2.0 open weights; tiers such as Medium 3.5 use a bespoke license
Pricing

Le Chat comes in Free, Pro, Team, and Enterprise tiers; the free tier is usable and advanced features need a subscription (check the official page for amounts). The API on La Plateforme bills per token — Mistral published $0.50 per million input and $1.50 per million output tokens for Mistral Large 3, and $0.15 / $0.60 for Mistral Small 4. Apache 2.0 open weights are free to self-host; you only pay for compute.

Pricing changes over time; check the official site

Mistral AI is one of Europe’s best-known large language model companies, headquartered in Paris, France. Amidst a landscape dominated by closed-source giants, it has carved out a distinct path: releasing high-quality models with open weights while attracting developers through cost-effective APIs and open licensing. It is widely regarded as "the European challenger to OpenAI."

What is Mistral AI?

Mistral offers both models and end-user assistants:

  • Open-weight models: From the early Mistral 7B and Mixtral through Mistral Large 3, released December 2, 2025, the weights ship under Apache 2.0 for free download, fine-tuning, and commercial use. Large 3 is a sparse MoE with 675B total and 41B active parameters, a 256K context, and text-plus-image input — Mistral's first MoE since Mixtral.
  • Mistral Small 4: released March 16, 2026, it folds the previously separate Magistral (reasoning), Pixtral (multimodal), and Devstral (agentic coding) into one model — 128 experts with 6B active parameters per token out of 119B total, also Apache 2.0, with a reasoning_effort setting to trade speed against depth.
  • Bespoke-license tiers: models such as Medium 3.5 use a modified license rather than Apache 2.0, so check the terms when choosing a tier.
  • Le Chat: The official conversational assistant, accessible via web and app, supporting web browsing, document processing, image understanding, and more.
  • La Plateforme: A developer API platform billed per token.

Why Developers Like It

  • Cost-effective: For comparable capabilities, API prices are typically lower than those of leading closed-source models, making it suitable for cost-sensitive, large-scale deployments.
  • Self-deployable: Open-weight models can run on your own servers, meeting data compliance and privacy requirements—a critical factor for organizations in Europe and heavily regulated industries.
  • Fast response and small footprint: Some models emphasize lightweight efficiency, performing well in latency-sensitive scenarios.

Getting Started: From Le Chat to the API

  1. Try it in Le Chat first: sign in on the web or in the app and chat to get a feel for the model's quality and speed; the free tier is enough.
  2. Get an API key: sign up on La Plateforme, add a payment method and create a key. The API and a Le Chat subscription are billed separately — Pro doesn't include API usage.
  3. Get the pipeline working on a small model: use an inexpensive tier such as Small 4 to get calls, parsing and error handling working, then move to a stronger model as quality requires. Small 4's reasoning_effort setting switches between quick answers and deeper reasoning, so tune it per task rather than always running the highest level.
  4. Check hardware before self-hosting: full open-weight versions are large — Small 4's BF16 weights are about 238GB, and even quantized versions need 60GB or more of GPU memory — so they suit servers rather than personal computers.

Who Should Use It

  • Developers and startup teams looking to embed large language models into applications while controlling costs.
  • Organizations with strict data sovereignty and compliance requirements (e.g., in Europe, finance, or government sectors).
  • Technical teams wanting to run open-source models locally or on private clouds for industry-specific fine-tuning.

Usage Recommendations

For everyday conversations, simply use Le Chat. To integrate into a product, start by testing workflows with smaller models on La Plateforme, then upgrade to stronger tiers as needed based on quality requirements, balancing cost and performance. Open-weight models are ideal for scenarios requiring long-term private deployment and avoiding vendor lock-in.

How It Compares

  • Other open-weight options with stronger Chinese: Qwen and DeepSeek.
  • If you just want one endpoint to switch between providers and compare prices: Mistral's models are also available through OpenRouter.
  • For Chinese-first products: Mistral's Chinese is usually weaker than Chinese models', so compare on your own real samples before launch.

Pricing

Le Chat and the API bill separately — a Pro subscription does not cover API calls.

  • Le Chat: Free, Pro, Team, and Enterprise tiers. The free tier is usable; advanced features and higher message limits need a subscription. Check the official pricing page for current amounts.
  • API (La Plateforme): billed per token. Mistral published $0.50 per million input and $1.50 per million output tokens for Mistral Large 3, and $0.15 / $0.60 for Mistral Small 4.
  • Open weights: Apache 2.0 tiers are free to download and self-host; you only pay for compute. Note that Small 4's full BF16 weights run about 238 GB, dropping to roughly 60–70 GB with 4-bit quantization.