Mistral

Latest AI news, models and releases from Mistral. ['AB-UPT', 'CodeStral-22B', 'Connectors', 'Devstral', 'Devstral 2', 'Devstral 2 Small', 'Devstral Small 2', 'Document AI', 'Emmi AI', 'Forge', 'GyroSwin', 'Leanstral', 'Leanstral 1.5', 'Leanstral (предыдущая версия)', 'Le Chat', 'Le Chat (Vibe)', 'Magistral', 'MiMo-V2-Flash', 'Ministral 3', 'Ministral-3-3B-Instruct-2512', 'Ministral-3-8B-Instruct-2512', 'Ministral 3B', 'Mistral', 'Mistral 3', 'Mistral 7B', 'Mistral-7B', 'Mistral AI', 'Mistral Large', 'Mistral Large 2', 'Mistral Large 3', 'Mistral-Large-V2', 'Mistral Medium 3.1', 'Mistral Medium 3.5', 'Mistral Nemo', 'Mistral OCR 3', 'Mistral OCR 4', 'Mistral OCR4', 'Mistral Search Toolkit', 'Mistral Small', 'mistral-small-2603']

Official site ↗

Open Source 🇺🇸 ⚡

Gigatoken: BPE Tokenizer in Rust Encodes Text at 24.53 GB/s, Up to 989x Faster Than HuggingFace Tokenizers

Gigatoken, a BPE tokenizer written in Rust by Stanford PhD student Marcel Rød, achieves text encoding speeds of 24.53 GB/s on a 144-core AMD EPYC system, outperforming HuggingFace tokenizers by 989x and OpenAI's tiktoken by 681x. The library supports 23 tokenizer families and attributes its speed to a hand-optimized pretokenizer using SWAR SIMD techniques and pretoken caching.

OpenAI

Hugging Face

Google/DeepMind

Top Open ASR Models of 2026: WER, Languages, Latency, and License Comparison

The open speech recognition landscape has diversified beyond Whisper, with multiple models achieving sub-6% WER on the Open ASR Leaderboard. Key developments include Cohere Transcribe, IBM Granite Speech 4.1, and various models for throughput, streaming, and language coverage. The report emphasizes that license, language support, throughput, and streaming capability are now more critical than raw leaderboard rank.

Cohere

IBM

NVIDIA

Alibaba/Qwen

Mistral

Kyutai