NVIDIA Nemotron 3.5 Transcribe leads open-weight ASR models at roughly 12 cents per hour

What NVIDIA shipped

NVIDIA released Nemotron 3.5 Transcribe (full name NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) on October 28, 2026, according to two dev.to guides published the following day. It is an open-weight speech-to-text model built on a hybrid Mamba-2 and Transformer mixture-of-experts language model, with a 600-million-parameter ASR encoder and a speaker encoder attached. The system totals 34 billion parameters but only about 4.3 billion are active per token, which is why it fits on a single GPU. The weights ship under the OpenMDW-1.1 license, which permits commercial use.

From a single checkpoint and a single call, the model returns transcription in 40 languages, word-level timestamps via a CTC alignment head, segment timings, speaker attribution for up to eight speakers, translation to and from English across 31 languages, detected language, and English-only audio question answering. It accepts recorded files up to one hour long within a 65,536-token context. One caveat repeated in both guides: this is an offline model for recorded files, not live microphone streams.

Accuracy and leaderboard position

On the Hugging Face Open ASR Leaderboard, updated the day of release, the model scores an average word error rate of 5.33%, ranking ninth out of 61 entries. Every model above it — ElevenLabs Scribe v2, Azure Speech and Zoom Scribe v2 Pro among them, per the dev.to write-ups — is a closed commercial API. Among models with downloadable weights, nothing ranks higher. OpenAI's whisper-large-v3 sits at position 41 with 7.16% WER.

NVIDIA's own model card reports a better-sounding 4.71% average, but the dev.to guide notes that the leaderboard figure folds in two private test sets alongside nine public ones, so both numbers are honest while measuring different mixes.

Price and access

Hosted inference is cheapest on DeepInfra at $0.002 per minute — about 12 cents per hour of audio, or roughly $120 for 1,000 hours. The dev.to API guide contrasts that with OpenAI's Whisper API at $0.006 per minute, three times the rate. OpenRouter also lists the model at $0.000033 per second, but as of October 29, 2026, DeepInfra was its only serving provider, so there is no failover path. A dense hour of speech can generate around 20,000 output tokens, so clients should set long timeouts.

There are also three free routes. NVIDIA runs a public demo Space on Hugging Face with karaoke, podcast and movie-subtitle tabs and no login required. Community Spaces wrap the same checkpoint. Most substantially, self-hosting costs nothing beyond hardware: the reference BF16 setup needs an H100 80GB (H200, A100 80GB, B200, GB200 and DGX Spark also qualify), FP8 adds the L40S, and an NVFP4 checkpoint runs on a consumer RTX 5090 or RTX PRO 6000 Blackwell. Older 24 GB cards such as the RTX 4090 are not on the compatibility list. NVIDIA claims one RTX PRO 6000 Blackwell can process 100 hours of audio in under three minutes with the NVFP4 build, though the dev.to guide cautions that figure relies on serving optimizations that are not yet open source. For production self-hosting, the guides point to vLLM 0.32.0 with the NeMo Speech plugin.

Promptable transcription

Unlike classic ASR APIs, the model accepts instructions in plain language: verbatim wording with fillers, cleaned-up readable text, stable speaker-turn tags, or translation, plus context terms for names, products and jargon. NVIDIA reports that context prompting cuts WER on its ContextASR-Bench from 3.33% to 1.71%, and on the Earnings22 set from 12.54% to 9.58% — for earnings calls, the suggested approach is passing company names, tickers and executive names alongside the audio.

Why it matters

The top of the Open ASR leaderboard remains closed commercial territory, but the best downloadable model now sits inside the overall top ten, narrowing a gap that long pushed serious transcription workloads toward paid APIs. For developers, this changes the economics: word-level timestamps, eight-speaker diarization and 40-language transcription arrive at a third of OpenAI's hosted Whisper pricing, or essentially for free on owned hardware under a license that allows commercial embedding. The trade-off to watch is supply concentration. With a single hosted provider and full-precision hardware requirements starting at an H100-class GPU, anyone without modern accelerator hardware is betting that DeepInfra's twelve-cents-per-hour rate holds.