NVIDIA open-sources Nemotron 3.5 Transcribe, top open ASR in 40 langs
It beats Qwen3-ASR on the open leaderboard and ships with Granary v2, a 463K-hour open speech dataset — plus Liquid AI's six 2B edge models and a 1.3B OCR drop.
Copy markdown
One open model for transcribe, translate, diarize
Nemotron 3.5 Transcribe is a 34B / 4.3B-active hybrid-MoE that does verbatim or cleaned transcription, speech translation, speaker diarization, word timestamps and audio Q&A across 40 languages. It tops the English Open ASR Leaderboard at 4.71 mean WER (vs Qwen3-ASR's 4.94) and runs ~410x realtime on a single H100 — roughly 100 hours of audio in under three minutes on Blackwell. The OpenMDW-1.1 license clears commercial use, and you serve it behind an OpenAI-compatible vLLM endpoint.
Granary v2: 463K hours to fine-tune your own ASR
NVIDIA shipped the model alongside Granary v2, an open multilingual speech corpus of ~463,000 hours. That's the real builder unlock: NVIDIA's own medical-speech fine-tune cut WER from 30.8% to 11.8%, so you can adapt one checkpoint to your domain and languages instead of stitching together Whisper, a separate diarizer and a translator.
Liquid AI's LFM3: six 2B models for the edge
Liquid AI open-weighted six ~2B models on one backbone — text, vision, audio, and an any-to-any omni — tuned to run on phones, a Raspberry Pi 5 and laptops. The text model fits in about 1 GB at 4-bit with a 256K context, the omni beats Qwen2.5-Omni-7B on audio-video benchmarks, and you can fine-tune any of them on a single GPU via llama.cpp, MLX or vLLM.
A 1.3B OCR you can run in Ollama
OpenBMB open-sourced MiniCPM-OCR-1.0, a 1.3B document-OCR model aimed at scanned pages. It's already on Ollama and Hugging Face, so a local, no-API document-to-text pipeline now fits on modest hardware — a lightweight counterpart to the heavier Youtu and Baidu parsers that dropped earlier today.