Nemotron 3.5 Transcribe vs Voxtral 3: The Two-Point Gap That Doesn't Settle It

One model, three word error rates. Artificial Analysis lists NVIDIA Nemotron 3.5 Transcribe at 3.4% on its non-streaming leaderboard; the same page's FAQ describes it at 2.8%. Meanwhile Mistral Voxtral 3 — the family that ships as Voxtral Small Transcribe 3 — sits at 2.1%, third on that board and roughly half a point clear of NVIDIA's best published figure. If you take the 3.4%, the gap looks decisive. If you take the 2.8%, it is inside the noise band this index has shown across serving endpoints. And if you look at the sub-score that covers conversational audio, the ranking reverses.

Two numbers for one model, and neither is a lie

The 3.4% and the 2.8% come from the same leaderboard on the same day. They differ because AA-WER v2 is computed against a specific evaluation slice through a specific serving endpoint, and the index has historically moved by close to a full point between providers running the same checkpoint. NVIDIA's 2.8% is the number the company quotes in its launch material; the 3.4% is what the board shows for the endpoint it actually tested. Neither is fabricated, and neither should be subtracted from Mistral's 2.1% as if the two were measured on identical hardware, at identical precision, against identical audio.

That does not make the gap meaningless. It makes it a range: somewhere between 0.6 and 1.3 points in Voxtral's favour on the batch index, with the true value depending on whose endpoint you use. For a customer-service deployment where every point of WER costs human review minutes, that is worth money. For a meeting-notes product where a human skims the output anyway, it is not.

Where the difference actually comes from

The AA index is an average over three components: VoxPopuli, Earnings-22 and AgentTalk. The first two are read speech and earnings calls — clean, single-speaker, well-recorded, the kind of audio both models handle well. AgentTalk is conversational and noisier. On that slice, NVIDIA's model reportedly scores 1.8% against Voxtral Small's 2.3%, which is a real reversal and the most interesting number in the comparison.

• VoxPopuli and Earnings-22 — Voxtral Small leads; these corpora reward a model tuned on clean long-form speech.

• AgentTalk — Nemotron leads, 1.8% against 2.3%, on the conversational audio that resembles an actual support call or a meeting with crosstalk.

• Overall index — Voxtral Small at 2.1%, Nemotron between 2.8% and 3.4%. The clean-speech components carry enough weight to produce that order.

That pattern is consistent with the two architectures. Voxtral Small is a transcription model trained to be excellent at transcribing; NVIDIA's checkpoint is a language model with a speech encoder bolted to the front, and language models tend to do better when the audio is messy and the context is conversational. It also explains why NVIDIA's biggest claimed wins are not on clean WER at all: 15.80 average cpWER on speaker-attributed overlapping speech, and a named-entity WER that falls from 12.21% to 2.20% when you supply an entity list.

The vendor tables do not overlap either

Mistral publishes FLEURS-45 macro-average WER: 4.1% forced-language and 4.0% inferred for Voxtral Small Transcribe 3, against 4.6% for its 4B Mini sibling and 16.0% for Whisper large-v3. NVIDIA publishes the Open ASR Leaderboard: 4.71% mean across nine English sets and 4.62% across seven multilingual sets. Both are vendor-run harnesses on different corpora with different normalisation, and the near-identical values are a coincidence, not a head-to-head. Anyone presenting 4.0 against 4.62 as a 0.6-point win is comparing a FLEURS score to an Open ASR score.

The diarization numbers have the same problem in reverse. Mistral reports 18.4% concatenated minimum-permutation WER on AMI, Switchboard, CallHome and AISHELL-4. NVIDIA reports 15.80 average cpWER on its own multi-speaker set. NVIDIA's number is lower and its set is different; without a shared corpus, the only defensible statement is that both companies believe they beat the specialised cascades they tested against, and neither has published a run on the other's data.

What Voxtral 3 does that Nemotron cannot

Mistral's release is a family, not a checkpoint, and the family covers ground NVIDIA's single offline model does not.

• Language breadth — Voxtral Small Transcribe 3 covers 50+ languages including all 24 official EU languages; NVIDIA covers 40, with speaker labels on 13.

• Real-time — Voxtral Mini Transcribe Realtime 3 streams at sub-500ms latency for $0.008 per minute, with 7.4% WER at a 240ms delay and 5.7% at 2,400ms. NVIDIA has no streaming mode for this checkpoint at any price.

• Voice agents — the Voxtral 3 launch shipped a LiveKit-based Voice Agent SDK that chains transcription, a reasoning model with tool calls, and speech generation, with barge-in and SIP telephony. That is an entire product category NVIDIA's checkpoint does not address.

• Operational simplicity — per-minute pricing, no GPU, no serving stack, EU, US and global endpoints. For a team without infrastructure, that is the difference between a two-day integration and a two-month one.

• Context biasing — 94.6% keyphrase recall with context against 81.4% without, on Earnings-22. Comparable in spirit to NVIDIA's entity-list mechanism, measured on a different corpus.

What Nemotron does that Voxtral 3 cannot

The reverse list is shorter but heavier for the right buyer.

• Open weights — OpenMDW-1.1, commercial use permitted, published in BF16, FP8, NVFP4 and GGUF. Mistral's page offers API access and "self-deployment on your own infrastructure" through Mistral, and as of publication publishes no weights, no licence and no checkpoint. If your compliance team needs to audit the artefact, only one of these exists to audit.

• One-hour single pass — 65,536 tokens of context and recordings up to an hour. Mistral does not publish a maximum recording length or context window for Voxtral 3, which is itself a reason to test before committing a long-form workflow.

• Translation — 31 languages in both directions to and from English. Voxtral 3's launch page documents transcription, diarization and context biasing; it does not document a speech-translation task the way NVIDIA's checkpoint does.

• Promptable output — five transcript styles selected by instruction, rather than one format plus post-processing.

• Audio understanding — question answering and spoken instruction following over the recording. English-only on this checkpoint, and not documented anywhere in the Voxtral 3 launch materials.

The cost comparison is not close, and that is the point

Voxtral Small Transcribe 3 is $0.0040 per minute, or $0.24 per audio hour. The 4B Mini is $0.0035, or $0.21 per hour. NVIDIA's hosted listing is $2.00 per 1,000 minutes, or $0.12 per hour. Self-hosting is a different arithmetic entirely: at NVIDIA's claimed 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell, a thousand hours of audio — 3.6 million seconds — takes about 27 minutes of GPU time. Even at half that throughput, which is what NVIDIA says the open-source integration delivers, it is under an hour.

• 1,000 hours per month — Voxtral Small $240, Voxtral Mini $210, Nemotron hosted $120, Nemotron self-hosted roughly $2 of GPU time at a $2/hour rental rate, plus operations.

• 10,000 hours per month — $2,400, $2,100, $1,200, and about $20 of GPU time respectively.

• The catch — the self-hosted line assumes you already have the GPU, the serving stack and someone who can keep it running. A single RTX PRO 6000 is a five-figure capital purchase; at 1,000 hours a month you will not amortise it before the next generation ships. At 10,000 hours it pays back in months.

• The other catch — those throughput figures depend on NVFP4 plus the DFlash2 draft model, and that combination disables word timestamps. If you need alignment, you are running a slower configuration than the one in the marketing table.

So the pricing argument splits by volume and by whether the audio may leave your network. Mistral's per-minute rate is cheap enough that most teams under a few hundred hours a month should not think about GPUs at all. Above ten thousand hours, self-hosting an open checkpoint becomes the cheaper option by an order of magnitude, and the operational burden is the only real question.

The fifty-hour test

Every number above is somebody else's audio. Before committing either model to a production path, transcribe fifty hours of your own recordings with both and score them the way your users will: named entities correct, speaker turns correct, numbers and currency amounts correct, and the specific accents your customers have. The AA index will not tell you which model handles a Glasgow call centre better, and neither vendor's harness will either.

That test is also where a router is useful for something other than cost. OrcaRouter puts the downstream text calls — summarisation, entity extraction, redaction, translation of the finished transcript — behind one key with 0% markup, provider list price passed through, automatic failover across providers, and a routing DSL that can send different stages to different models in one workflow. Neither Nemotron 3.5 Transcribe nor Voxtral 3 is in OrcaRouter's catalogue, so the router does not help you call them; it helps you process what they return without standing up a second vendor relationship for the text layer.

The verdict depends on which sentence describes your problem. If it is "we need accurate transcription in fifty languages, we do not want to own a GPU, and latency matters," Voxtral Small Transcribe 3 is the better product and the 2.1% is a real reason. If it is "we need speaker-attributed transcripts of hour-long meetings, translation, and an artefact our compliance team can inspect," Nemotron 3.5 Transcribe is the only option on this list that meets all three, and a point of WER on clean read speech is the price of that. If it is "we need both," the honest answer is that they are not substitutes, and the teams getting the most out of this generation are running one for live audio and the other for the archive.