Build my stack
Type
Major Release
Primary Source
Mistral AI Announcement
Published
Jul 24, 2026
Key Takeaways
- 1.1T total params, 84B active per token via sparse MoE.
- Apache 2.0 open weights for self-hosted deployment.
- Outperforms Llama 4, matches GPT-5.5 on reasoning.
- Mistral Vision: end-to-end multimodal encoder.
Paris-based Mistral AI has released Mistral Large 3, featuring 1.1 trillion parameters with 84 billion active per forward pass via sparse MoE. Ships under Apache 2.0 open weights.
Architecture and Capabilities
Sparse MoE activates only 84B of 1.1T params per token. 128k context window. Mistral Vision handles schematics, PDFs, and data visualizations natively.
Enterprise Deployment
Self-hosted: 8x H200 GPUs at FP8. API: $2.50/1M input, $7.50/1M output tokens via La Plateforme.
EU AI Act Compliance
Machine-readable provenance metadata, safety alignment cards, zero-data-retention modes included.
Mistral Large 3Mistral AI 2026open weights model1.1 trillion parameters
Sources & References
Frequently Asked Questions
Hardware required?
8x H200 GPUs at FP8, or 8x A100 80GB with 4-bit AWQ quantization.
API pricing?
$2.50/1M input tokens, $7.50/1M output tokens.
About the author
MV
Marcus VanceAI Hardware & Model Specialist
More Major Release coverage
All stories
Major Release
vlt 1.0 Ships with Hosted JavaScript Package Registries and Ecosystem Mirrors
Aug 4, 2026
Major Release
JetBrains Opens IntelliJ IDEA's Java and Kotlin Intelligence to VS Code, Cursor, and AI Agents
Aug 4, 2026
Major Release