Nvidia cut its own best transcription error rate from 4.3% to 2.8% and now shares the top open-weight spot, a reminder that its moat is software as much as silicon.
Nemotron 3.5 Transcribe's 2.8% WER on the AA-WER Index ties it with Voxtral Small for #1 among open-weight models, ahead of Thinking Machines' Inkling at 3.5%. It improves on NVIDIA's previous most accurate model, Canary Qwen 2.5B, at 4.3%. https://t.co/Hb5R76kipp
- Replies1
- Reposts0
- Likes6