All Organizations
Mistral AI

Mistral AI

News, updates, and in-depth coverage

10 Updates
Articles

Mistral Forge introduced

Mistral AI announced Forge, a new product in its platform lineup, expanding the Mistral ecosystem beyond models into developer tooling.

Mistral AI partners with NVIDIA to accelerate open frontier models

Mistral AI announced a partnership with NVIDIA to accelerate open frontier model development, with Mistral 3 models optimized across NVIDIA supercomputing and edge platforms.

Mistral Small 4: 119B MoE with 6B active parameters, Apache 2.0

Mistral AI released Small 4 — a 119B total / 6B active MoE model with 128 experts, 256K context, and configurable reasoning effort. Delivers 40% faster completion times and 3x throughput over Small 3. Apache 2.0 licensed.

Leanstral: open-source code agent for Lean 4 formal proofs

Mistral AI released Leanstral, the first open-source code agent for Lean 4 formal proof verification. With 6B active parameters and Apache 2.0 license, it achieves 26.3 at pass@2 for $36 vs Claude Sonnet at $549.

Voxtral Transcribe 2: sub-200ms real-time speech-to-text in 13 languages

Mistral AI released Voxtral Mini Transcribe V2 (batch, 4% WER, $0.003/min) and Voxtral Realtime (live, sub-200ms latency, 4B params, Apache 2.0). Both support speaker diarization, context biasing, and word-level timestamps.

Mistral OCR 3 released: 74% win rate over OCR 2

Mistral AI launched OCR 3 with major improvements in handwriting recognition, form detection, scanned documents, and complex table reconstruction. Priced at $2 per 1,000 pages ($1 with Batch API).

Devstral 2 and Mistral Vibe CLI launched for code agents

Mistral AI released Devstral 2 (123B, 72.2% SWE-bench Verified, modified MIT) and Devstral Small 2 (24B, 68.0%, Apache 2.0) alongside Mistral Vibe, an open-source CLI coding agent. Devstral 2 claims 7x more cost-efficient than Claude Sonnet.

Mistral 3 released: Large 3 (675B MoE) and Ministral 3 family

Mistral AI launched Mistral Large 3 — a 675B total / 41B active parameter MoE model trained on 3,000 H200 GPUs — alongside Ministral 3 in 3B, 8B, and 14B sizes. All Apache 2.0 licensed with native image understanding and 40+ languages.

Mistral AI Studio launched

Mistral AI introduced AI Studio, a platform for testing models, parsing documents with Document AI, and building AI-powered applications.

Mistral AI raises 1.7 billion euros at 11.7B euro valuation

Mistral AI closed a Series C round of 1.7 billion euros led by ASML (1.3B euro, 11% stake), with participation from DST Global, Andreessen Horowitz, Bpifrance, General Catalyst, Lightspeed, and NVIDIA. Became France's first decacorn.

You've reached the beginning