LFM3-Audio-2B vs Voxtral 3: A 2.1 GB Download Against a Voice Platform

The honest way to open a comparison between LFM3-Audio-2B and Voxtral 3 is to say that nobody has published a run of it. Liquid AI published a benchmark table for its 2-billion-parameter audio model, and Mistral published one for its speech suite, and the two tables share almost no rows. Where they do touch, the models being measured are not the same generation. Anyone telling you which of these is "better" on a benchmark is comparing a number from one test harness to a number from another and hoping you do not check.

So this piece does something less tidy and more useful. It separates the question you can answer from the one you cannot. You cannot yet say which model transcribes German better, because there is no shared multilingual test between them. You can say, with confidence and from the two vendors' own documentation, that these are not the same kind of product at all: one is a 2.1 GB file you run yourself, and the other is a four-model pipeline with a telephony SDK attached and no downloadable weights.

That difference decides most real deployments before a single WER figure is consulted. The rest of this article is about where it does not.

The Voxtral Liquid benchmarked is not the Voxtral you are comparing against

Liquid's model card puts LFM3-Audio-2B next to six competitors, and one of them is a Voxtral. The row is labelled Voxtral Mini 3B, and on English ASR it beats Liquid's model: 5.51% word error rate against 6.23%. That is the one direct Voxtral comparison Liquid published, and it is easy to read as "Mistral's audio model transcribes more accurately."

It is not the comparison the headline implies. Voxtral Mini 3B is mistralai/Voxtral-Mini-3B-2507 — a 3B model released in July 2025 under Apache 2.0, built for local and edge deployment. It predates Voxtral 3 by fifteen months and belongs to a different product generation. Mistral's October 20, 2026 suite contains no model called Voxtral Mini 3B; its smallest transcriber is Voxtral Mini Transcribe 3 at 4B parameters and $0.0035 per minute.

There is also a discrepancy inside the table. The parameter column on that row reads 5B, on a row named for a 3B model. That is not a reason to distrust the WER figures, which come from a named evaluation — ESB and the Open ASR leaderboard set, averaged. It is a reason to read the column headings on vendor tables before you read the numbers.

The consequence is worth stating plainly, because it is the single most useful fact in this article: there is no published head-to-head between LFM3-Audio-2B and any Voxtral 3 model. Not from Liquid, not from Mistral, and not from an independent evaluator. Liquid compared its new model to Mistral's open-weights audio model from 2025, which is a defensible choice for a fair open-versus-open comparison — and is also the only kind of comparison now possible, because Voxtral 3 ships no weights. The 2026 open-weight audio comparison cannot be run against Voxtral 3 at all.

What the two products actually are

The asymmetry is easier to see in the shape of the release than in the specs.

LFM3-Audio-2B is one model, released October 27, 2026 as part of the six-model LFM3 family. It takes text or 16 kHz audio in, and puts out text, 24 kHz speech, or interleaved text and speech, plus tool calls. A new in-house causal audio encoder lets it prefill while you are still speaking. Its LiquidCodec tokenizer compresses 80 milliseconds of speech into 12 discrete tokens, decoded in parallel rather than sequentially, at roughly 87 times real time on a single CPU thread of an M4 Max — a codec throughput figure, not a hardware requirement.

Voxtral 3, announced October 20, 2026, is a suite: a batch transcriber, a second smaller batch transcriber, a streaming transcriber, a TTS model, a voice-adaptation product, and a LiveKit-based voice-agent SDK that stitches them together over SIP and WebRTC. You do not pick Voxtral 3. You pick a pipeline, and Voxtral 3 supplies the stages.

• Weights — LFM3-Audio-2B ships open weights under the LFM Open License v1.0; Voxtral 3 publishes no downloadable weights, and Mistral's launch material describes API access, Studio playgrounds and private infrastructure deployment instead.

• Footprint — 2.1 GB on Ollama against an endpoint; there is no Voxtral 3 file to size.

• Context — 32K tokens for LFM3-Audio-2B, input plus output. Mistral publishes no context figure for the transcribers at all; it prices by the minute instead.

• Modalities — LFM3-Audio-2B is any-to-any in one model: hear, speak, understand, translate, call tools. Voxtral 3 splits those across separate models and endpoints; the transcribers and the TTS model are not the same system.

• Licence — LFM Open License v1.0, free for commercial use below $10 million in annual revenue; Voxtral 3 is a commercial API whose terms apply per call.

• Diarization and context biasing — Voxtral 3 has both, with an 18.4% cpWER on Mistral's own diarization set and keyphrase recall rising from 81.4% to 94.6% when you supply context. LFM3-Audio-2B publishes neither capability.

Language coverage is the one gap that is not close

LFM3-Audio-2B supports 18 languages across 19 locales, listed on the card: Arabic, Danish, German, Greek, English (British and American counted separately), Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish and Chinese. It benchmarks multilingual ASR on nine of them and reports a 6.03 mean error rate on FLEURS, using character error rate for Chinese, Japanese and Korean.

Voxtral 3's transcribers cover 50 or more languages. The realtime model claims 35+ in one place on the page and 40+ in another, which is the kind of inconsistency that usually means the number moved during launch week and nobody reconciled the copy.

If your audio arrives in Vietnamese, Hindi, Russian, Norwegian or Catalan, LFM3-Audio-2B does not list it, and Voxtral 3 speaks all five — they are among the eighteen voice-preset languages Mistral names. Across the whole list, we could not identify a language Liquid supports that Mistral's 50+ transcription languages do not claim to cover. The honest caveat is that Mistral publishes its voice-preset list in full and its transcription coverage only as a count, so part of this comparison is a list against a number.

There is a sharper edge inside Liquid's own documentation. The model's live ASR mode — the one that transcribes as audio streams in, in chunks of about 0.96 seconds — supports English only. Batch transcription covers the full 18-language list; the streaming path does not. Anyone planning a realtime multilingual captioning feature off LFM3-Audio-2B should read that line twice, because it rules the model out for the job regardless of the WER table.

Two latency numbers that do not measure the same thing

Liquid advertises time to first token "as low as 50 ms." Mistral advertises sub-500 ms for its realtime transcriber and a full agent turn under one second. Those numbers are not comparable, and the difference is not a rounding error.

The 50 ms figure is a speech-to-speech path: the causal encoder begins processing before you finish the sentence, so the model can start emitting the first token of its reply while audio is still arriving. It is a claim about interactivity in a conversational model. The sub-500 ms figure is transcription latency — the delay before text appears — and Mistral's own tables show what it costs. On its realtime benchmark, Voxtral Mini Transcribe Realtime 3 at 960 ms of latency reads 6.0% WER; at 480 ms it is 6.4%; at 240 ms it is 7.4%. Latency and accuracy trade directly against each other, and Mistral lets you pick the point on that curve.

Liquid publishes no equivalent curve. It publishes a floor — 50 ms — with no stated hardware, no test set, and no accuracy at that latency. The floor may be real. It is not yet a specification you can plan a product against.

The cost math, worked rather than asserted

Comparing an open-weights model to a per-minute API comes down to volume, and the arithmetic is simple enough to do properly.

Transcription, 1,000 hours of audio — 60,000 minutes. Voxtral Mini Transcribe 3 at $0.0035/minute is $210. Voxtral Small Transcribe 3 at $0.004 is $240. Voxtral Mini Transcribe Realtime 3 at $0.008 is $480. LFM3-Audio-2B is a free download running on hardware you already own, so its marginal cost is electricity.

Speech synthesis, 1,000 hours — at roughly 900 characters per spoken minute, that is 54 million characters. Voxtral Mini TTS 3 at $13 per million characters is $702. LFM3-Audio-2B costs nothing per character, and offers two voice options, male and female, selected by prompt. Voxtral clones a voice from 30 minutes of audio and exposes emotion, pacing and emphasis controls.

A live voice agent, 1,000 hours of conversation — Mistral prices the assembled pipeline at $0.04 per minute with Mistral Small 4 doing the reasoning, $0.08 with a larger model. That is $2,400 to $4,800. LFM3-Audio-2B handles the hearing and speaking locally, but it still needs a reasoning model for anything beyond its own competence, and its card says outright that it is not recommended for complex multi-step reasoning.

Now the part usually left out: break-even. A $1,000 machine pays for itself after about 4,760 hours of Voxtral Mini Transcribe 3 — roughly 200 days of continuous audio, or a year and a half at eight hours a day. At small volumes, the API is cheaper than buying hardware, and it is cheaper on the day you start. The open-weights case is not "free." It is "fixed cost instead of marginal cost," and it only wins when the volume is there, the privacy requirement is hard, or the device has no network.

Where a router earns its place in this picture is not on either audio model — neither LFM3-Audio-2B nor Voxtral 3 is on the OrcaRouter catalogue today, and we are not going to pretend otherwise. It is on the reasoning stage both architectures end up needing. A voice agent is a chain: transcribe, reason, speak. The middle link is a hosted LLM, and that is the link that benefits from one key across more than 200 models, provider list prices passed through at 0% markup so a vendor price change reaches you the same day, and automatic failover when a provider has a bad hour mid-call.

What is verified, and what is not

Three tiers of evidence are in play here, and flattening them is how comparisons go wrong.

Independently measured. Voxtral 3's transcription quality has third-party numbers: Artificial Analysis placed Voxtral Small Transcribe 3 at 2.1% WER on its own index, third overall as of October 21, 2026, with Voxtral Mini Transcribe 3 at 2.2%. Its TTS landed at rank 13 on the Provider Voice arena with 1,217 Elo and rank 14 on Controlled Voice with 1,066 — mid-pack, and worth saying plainly given Mistral's social post described it as top twelve.

Vendor-reported, unreproduced. Every LFM3-Audio-2B figure in this article is Liquid's own: the 6.23 English WER, the 6.03 multilingual figure, the 3.83 UTMOSv2, the 73.13 VoiceBench, the 58.50 internal tool-calling eval and the 50.60 MMAR. There is no Artificial Analysis entry for the model as of this writing, and no third party has published a reproduction. Treat them as claims, not measurements.

Not comparable at all. Liquid's 6.23% English WER and Mistral's 4.0% FLEURS figure and Artificial Analysis's 2.1% are three different tests on three different audio sets. Do not subtract them. The only same-harness comparison Liquid published is against Voxtral Mini 3B — the 2025 open-weights model. There Liquid's 2B scores 6.23 against Voxtral's 5.51 on English ASR, but beats it on multilingual ASR 6.03 to 7.76, VoiceBench 73.13 to 65.78, MMAR 50.60 to 50.10 and tool calling 58.50 to 44.50.

What "shipped quietly" actually means here

LFM3-Audio-2B did not get a launch of its own, and that distinction matters if you are deciding how much to trust the documentation.

Liquid announced the LFM3 family on October 27, 2026 with a post titled "Introducing LFM3: Every Modality for Specialized Intelligence." The audio model appears there as one of six, in a section — there is no dedicated LFM3-Audio blog post, no separate launch, no product page of its own. The weights and the model card are complete and public. The Ollama listing, ollama run lfm3-audio at 2.1 GB, appeared on October 28 without a corresponding announcement from either Liquid or Ollama, alongside a llama.cpp pull request that added the audio and omni models with live ASR deferred to a follow-up.

So the framing is: the artifact is real and documented, the launch was a paragraph inside a family announcement, and the local-runtime support arrived a day later, quietly. What that means practically is that the model card is the authority — not a launch post, because there is not one — and the card is unusually specific about limitations, which is a point in its favour. It names LiveASR as English-only. It says the model is not for agentic coding or complex multi-step reasoning. It sends knowledge-heavy tasks to retrieval. A vendor hiding a weak model does not usually write that down.

Who should take which

Take LFM3-Audio-2B if the audio cannot leave the device, if you need one model that both hears and speaks without stitching two services together, if your languages are among its eighteen, and if your volume is high enough that a fixed hardware cost beats a per-minute bill. It is also the right choice for prototyping a voice interface with no procurement at all — 2.1 GB and an Ollama command. Accept that its English ASR is behind the last generation of open Voxtral, that streaming is English-only, and that no independent evaluator has checked any of its numbers.

Take Voxtral 3 if you are building a phone agent rather than a model feature. The suite exists to be assembled: transcriber, reasoning model, TTS, over SIP and WebRTC, with barge-in and end-of-turn detection already solved. Take it too if you need 50+ languages, diarization, context biasing for names and jargon, or voice cloning — four capabilities LFM3-Audio-2B does not offer in any form. The trade is that you cannot download it, you cannot inspect it, and your cost scales linearly with every minute you serve, forever.

Take both if the architecture allows it, and it often does. Run LFM3-Audio-2B on the device for wake-word, local dictation and the offline path; send the calls that need multilingual coverage, diarization, or a cloned voice to Voxtral 3. The two do not compete for the same slot in a stack. One is the model inside your product; the other is the vendor you call when the product is a phone line.

The comparison to watch for

Two things would settle this properly, and neither exists yet.

The first is an independent evaluation of LFM3-Audio-2B on the same harness Artificial Analysis uses for the Voxtral 3 models. That single number — its WER index on a shared test — would convert the most consequential claim in Liquid's table from vendor assertion to measurement, and would tell you whether a 2B model running on a laptop is genuinely within a point of a hosted 4B.

The second is whether Mistral ever ships Voxtral 3 weights. The company's 2025 audio models were open under Apache 2.0; the 2026 suite is not. If that reverses, the fair open-versus-open comparison Liquid chose to publish becomes a live one again, and this article's central caveat expires.

Until one of those happens, the defensible position is the unglamorous one. On capability breadth, Voxtral 3 wins by a distance that no benchmark dispute can close: fifty languages, diarization, cloning, telephony. On deployability, LFM3-Audio-2B wins by an equally large margin, because it is the only one of the two you can hold. Everything in between is a number from one vendor's table next to a number from another's, and the honest answer to "which is more accurate" is that nobody has measured it.