NVIDIA released Nemotron 3.5 ASR on 2026-06-04 as a compact streaming model for 40 language-locales, with latency steps starting at 80 ms chosen at inference. It suits voice-agent and captioning teams that want one self-hosted multilingual model instead of a deployment per language.
Can you use it?
- DeepInfra API: yes (Pay as you go on DeepInfra; the listing is marked zero retention). Priced per audio minute: $0.0002 per minute.
- Free plan: not published.
- Open weights: released (OpenMDW-1.1).
Checked 2 Nov against NVIDIA's own pages. Compare access across frontier models
What changed
- · Release Nemotron 3.5 ASR released on Hugging Face (nemotron-3.5-asr-streaming-0.6b-v1) source
Nemotron 3.5 ASR is NVIDIA's open 600M-parameter streaming speech recognition model. Its cache-aware FastConformer-RNNT design averages 8.84% FLEURS word error rate across 19 transcription-ready locales at the 1,120 ms chunk, and a single checkpoint selects the language by prompt, with optional automatic detection.
Provider: NVIDIA · Family: Nemotron 3.5
Input modalities: audio · Output: text
About Nemotron 3.5 ASR
Nemotron 3.5 ASR is a multilingual, streaming speech recognition model from NVIDIA, published on Hugging Face on 4 June 2026. It has 600 million parameters and uses a cache-aware FastConformer encoder with 24 layers and an RNN-T decoder. NVIDIA built it on its English-only Nemotron ASR Streaming 0.6B checkpoint and added language-ID prompt conditioning, so one set of weights transcribes 40 language-locales instead of needing a model per language.
The model keeps caches for every encoder self-attention and convolution layer, so each audio frame is processed once and never recomputed, unlike buffered streaming where chunks overlap. The chunk size is chosen at inference time from 80, 160, 320, 560 or 1,120 ms, with no retraining, which lets a team trade latency against accuracy per product. Output includes punctuation and capitalization. Input is mono audio plus a language tag, and with target_lang set to auto the model detects the spoken language and appends a language tag after the terminal punctuation.
Coverage is tiered, and the tiers matter more than the headline number. NVIDIA lists 19 transcription-ready locales (including English US and UK, Spanish, French, German, Japanese, Korean, Hindi and Arabic), 13 broad-coverage locales (including Polish, Swedish, Mandarin and Finnish) and 8 adaptation-ready locales (including Greek, Hebrew and Thai) that the tokenizer recognizes but that need fine-tuning on in-domain data before they transcribe properly. Only the first 32 produce usable transcription out of the box.
On the FLEURS test sets, NVIDIA reports an average word error rate of 10.38% across the 19 transcription-ready locales at the 80 ms chunk and 8.84% at 1,120 ms when the language is supplied. The 13 broad-coverage locales average 25.86% at 80 ms and 22.13% at 1,120 ms. Japanese, Korean and Mandarin are scored as character error rate. These are vendor-run figures, and NVIDIA notes that imperfect text normalization across 40 locales can inflate them.
The efficiency claim is the other selling point. On a single H100, NVIDIA reports about 240 concurrent real-time streams at the 80 ms setting against 14 for its buffered Parakeet RNNT 1.1B multilingual model, and 2,400 against 400 at 1,120 ms. That is NVIDIA's own comparison, run on its own hardware, so treat the ratio as a ceiling until someone reproduces it on their traffic.
Access is split between weights and hosting. The weights are free under OpenMDW-1.1 and load through NeMo 26.06 or Transformers, on Linux with Ampere, Hopper, Blackwell, Jetson and older NVIDIA GPUs. DeepInfra serves the model as an API priced per audio minute (the figure is in the pricing note). For English-only work NVIDIA itself recommends the separate Nemotron ASR Streaming English checkpoint, and the hosted Gemini 3.5 Transcribe is the closest closed alternative on Hokai. Teams that want a finished speech product rather than a model can look at Deepgram, AssemblyAI or Speechmatics, and voice-agent builders comparing stacks can start with Vapi, Retell AI and the guide to AI voice assistants.
Other NVIDIA models on Hokai include Nemotron 3.5 Lightning, Nemotron 3.5 Super VL, Nemotron 3.5 Bolt and Nemotron 3 Nano, none of which are speech models. Google's Gemini 3.5 Transcribe comes from Google.
Skip it if you need speaker labels, translation or audio question answering, because this is a plain transcriber. NVIDIA's separate Nemotron 3.5 Transcribe, a 30B-A4B model for offline recordings, covers those, and the two names are easy to confuse. Also skip it if you have no GPU and no hosted provider you trust, since the card documents NVIDIA hardware and Linux only.
Pricing
The weights are free under the open licence, so self-hosting costs the GPU you run them on. Hosted on DeepInfra at $0.0002 per audio minute (USD, read 2026-11-02, not per token). [Together AI](/hub/companies/together-ai)'s pricing page did not list this model on 2026-11-02. For a finished product instead of a model, compare [Deepgram](/hub/tools/deepgram) or [AssemblyAI](/hub/tools/assemblyai).
Key Features
- Language-ID prompting: One checkpoint transcribes many language-locales by taking a language tag, and target_lang=auto detects and tags the language.
- Cache-aware streaming: Encoder caches reuse hidden states so each frame is processed once, with no overlapping computation between chunks.
- Runtime latency control: Chunk size is set at inference time across five steps, so latency becomes a deployment setting and not a training job.
- Punctuation and capitalization: Output text carries punctuation and casing natively, so no separate restoration model is needed.
- Tiered language coverage: The card sorts locales into transcription-ready, broad-coverage and adaptation-ready tiers, so teams can see what works before testing.
Pros
- One small model replaces per-language deployments across every locale that works out of the box.
- Open weights allow on-premises use, with a hosted option on DeepInfra for teams without GPUs.
- Latency is a runtime setting, not a retraining job.
Cons
- Eight locales are not usable without fine-tuning.
- All published accuracy and throughput numbers are NVIDIA-run.
- Documented only for NVIDIA GPUs on Linux, with no diarization or translation.
Frequently Asked Questions
What does Nemotron 3.5 ASR actually cost?
The weights are free to download under OpenMDW-1.1, so self-hosting costs only your GPU time. DeepInfra lists the hosted model at $0.0002 per minute of audio, which is $0.012 per hour. Google's reference rate for Gemini 3.5 Transcribe works out to about $0.005 per minute, roughly 25 times higher at list price.
Is Nemotron 3.5 ASR better than Gemini 3.5 Transcribe?
The published numbers cannot settle it, because NVIDIA scores FLEURS across two locale groups while Google cites a single 5.5% streaming FLEURS figure and a 2.6% Artificial Analysis average. Pick Gemini 3.5 Transcribe if you want diarization, timestamps and disfluency cleanup in one call. Pick Nemotron if you need open weights, on-premises audio or the lowest per-minute price.
Can you use Nemotron 3.5 ASR commercially?
NVIDIA's Hugging Face card states the model is ready for commercial use and governed by OpenMDW-1.1. Some news coverage described the licence as the NVIDIA Open Model License, but the card and DeepInfra's listing both link OpenMDW-1.1, so read the licence text before you ship. The checkpoint is nvidia/nemotron-3.5-asr-streaming-0.6b.
Does Nemotron 3.5 ASR run on a CPU?
NVIDIA's card documents NVIDIA GPUs and Linux, with the NeMo 26.06 runtime, and says nothing about CPU inference for this model. The 'runs on CPU' claims in launch coverage belong to the English-only Nemotron ASR Streaming 0.6B model. Test your own latency before assuming the multilingual checkpoint behaves the same way.
Who should use Nemotron 3.5 ASR, and who should skip it?
It fits voice-agent and live-caption teams that want one self-hosted model across the 32 locales that work out of the box and need to tune latency per product. Skip it for English-only products, where NVIDIA points to its English checkpoint, and for jobs that need speaker labels or translation, where Gemini 3.5 Transcribe or NVIDIA's Transcribe model fit better.