llama.cpp Merges Support for NVIDIA's Nemotron 3.5 Transcribe: Offline ASR in 40 Languages
jacek2023 · reddit · 2026-11-03
llama.cpp merged PR #31471 adding support for NVIDIA's Nemotron 3.5 Transcribe, an open-weight, fully offline speech/audio transcription model:
- Transcribes in 40 languages, handles recordings up to 1 hour with a 65,536-token context
- Hybrid MoE architecture (30B-A4B) activating 4.3B parameters including the audio encoder
- Weights released in BF16, FP8, and NVFP4 on Hugging Face
For local/offline transcription, this puts a high-quality multilingual ASR model within reach of consumer hardware.
More from Infra
- How AI assistant Sauna runs 2 million databases on Turso, 90k created daily — glcst · 2026-11-03
- MoE leaderboard shows two spikes: ~300 teams hit 2.4x using Vjeux's public kernels — Vjeux · 2026-11-03
- AMD earnings preview: cross-reading supply chain calls points to server share gains — tengyanAI · 2026-11-03
- DeepSeek-V4.1-Flash GGUF quantized build trends on Hugging Face — ISTA-DASLab · 2026-11-03
- 4x RTX PRO 6000 local AI build done, next up maybe AMD R9700 — dee_hw · 2026-11-03
- Agent37 offers always-on agent sandboxes at $4.76/mo, cheaper than a Hetzner VPS — ayushtweetshere · 2026-11-03