LFM3-Audio-2B vs Dessa: One You Can Download, One You Can Only Hear

Liquid AI's LFM3-Audio-2B is a 2B-parameter checkpoint with an open licence, a Hugging Face repository and a 2.1 GB build you can pull onto a laptop. Block's Dessa is a full-duplex speech model announced on 23 October 2026 behind a waitlist, with no weights, no parameter count, no API and no published benchmark. Put them side by side and the usual comparison collapses before it starts: one of these can be run, fine-tuned and measured by anyone reading this, and the other can only be listened to. The useful question is not which sounds better. It is what each company chose to publish, and what that choice lets you do on Monday morning.

Both arrive with the same warning label, from opposite directions. LFM3-Audio-2B shipped on 27 October 2026 inside the six-model LFM3 family drop, and the audio branch got one paragraph of Liquid's family post plus a model card. There is no standalone "Introducing LFM3-Audio" announcement, no audio pricing entry on Liquid's site, and — as of this writing — no independent benchmark run of the model anywhere. Dessa has the opposite problem: a dozen curated conversation recordings and almost no specification. Neither company has given an outsider enough to reproduce a result.

What each one actually published

Strip the adjectives and the record is short. Everything below is from Liquid's model card, its family post, or Block's own announcement; nothing has been independently verified.

• What you receive — LFM3-Audio-2B: a BF16 checkpoint, GGUF and MLX 4-bit conversions, a tokenizer, and codec weights in a separate audio_tokenizer directory. Dessa: a waitlist form and a set of video clips.

• Size — LFM3-Audio-2B: 2B parameters, 2.1 GB as the Ollama build. Dessa: not disclosed. Block says it fits on a MacBook Pro M4 Max, which implies a model small enough for 128 GB of unified memory, but that is an inference, not a number.

• Input and output — LFM3-Audio-2B: text or 16 kHz audio in; text, 24 kHz speech, or interleaved text and speech out. Dessa: audio in, audio out, with no text intermediate at any point.

• Context — LFM3-Audio-2B: 32K tokens covering input and output. This is not the 256K figure repeated in some LFM3 coverage; that belongs to the text model, LFM3-1.7B. Dessa: not published.

• Licence — LFM3-Audio-2B: LFM Open License v1.0, free for commercial use below $10M annual revenue. Dessa: no weights, so no licence to read.

• Benchmark — LFM3-Audio-2B: a six-row vendor table. Dessa: none.

That last row is where the comparison gets interesting. Liquid published numbers and no independent witness. Block published witnesses — journalists who listened to the same clips — and no numbers. A reader trying to choose between them is choosing which kind of gap to tolerate.

The one advantage Dessa has that cannot be downloaded

Dessa is built for full duplex: it listens while it speaks. The model can be interrupted mid-sentence, can drop a "right" or "mm-hm" while you are still talking, and can adjust its timing to yours. Block's demos lean on this hard — laughter in the middle of its own sentence, a volume change on request, a film title finished for a speaker who forgot it. The Verge, listening to the same recordings, found the conversations "a little awkward and uncanny", which is a fair description of a system learning timing in public.

LFM3-Audio-2B does not claim that. It is request-and-response with streaming: a causal audio encoder lets it prefill while you speak, which is where Liquid's "as low as 50 ms" time-to-first-token figure comes from, and LiquidCodec streams speech back. Streaming a reply is not the same as handling barge-in. If your product needs a conversation that can be talked over, LFM3-Audio-2B is the wrong shape of model, and no amount of fine-tuning changes the interface.

What Dessa's demo cannot tell you is how often it behaves that way when nobody is selecting the clip. Block says the recordings are live, unsteered and subtitled afterwards, which rules out the crudest form of demo fraud. It does not rule out selection. A team that records many conversations and publishes the best of them has proved that good conversations exist, not that the model is reliable.

The numbers Liquid published, and what they are worth

Liquid's model card carries six figures, four with a predecessor value. All of them compare LFM3-Audio-2B with LFM2.5-Audio-1.5B, the model it replaces. No test set, harness or variance is named.

• English ASR, word error rate — 6.23 against 6.13. Flat. The only row that did not move, and a gap small enough to be sampling noise.

• Multilingual ASR, WER/CER — 6.03. No predecessor figure published.

• English TTS, UTMOSv2 — 3.83 against 3.22. A 0.6-point move, well outside noise, on a scale where 4.0 is usually treated as near-human.

• VoiceBench — 73.13 against 55.17.

• Audio understanding, MMAR — 50.60 against 25.85. Nearly double.

• Tool calling from speech, internal eval — 58.50. No predecessor figure.

Against the other models in Liquid's own table, this is a good small model rather than a leader: Voxtral Mini 3B posts 5.51 WER and NVIDIA Audex 2B posts 5.99, both better than 6.23; Qwen2.5-Omni 3B takes audio understanding at 54.41; Gemma 4 E4B takes tool calling at 72.40. The honest summary is competitive recognition, a clear step up in voice quality, and solid reasoning for 2B — not a record in any column.

Dessa has no equivalent row, and the absence is not a technicality. A speech-only model trained without text has no native interface for a JSON schema, a PDF or an API call. Block concedes the open questions on its own page: how efficiently a speech-only model learns facts and languages, and whether in-context learning survives without text pretraining. Those are the difference between a compelling conversation partner and a general assistant, and neither company has published the test that would settle it.

Licence, access, and who can build on what

LFM3-Audio-2B's licence is the practical reason it wins most procurement conversations before quality is discussed. The LFM Open License v1.0 allows commercial use below $10M annual revenue and requires a paid licence above it. Weights are on Hugging Face; llama.cpp merged support in PR #31257 on 27 October; Ollama lists a 2.1 GB build; Transformers needs 5.20.2 or newer with trust_remote_code. The custom-code requirement is the thing to pin — the model loads from its own module and reads codec weights from a separate directory, so an unpinned environment is a support ticket waiting to happen.

Dessa has no comparable path. Access is a waitlist plus early access for a small group of AI safety researchers who apply with an institutional email. Block's stated reasons — curiosity, human–AI interaction, safety — are consistent with keeping a voice that reads emotion inside a small room. The trade-off is that gated access and independent evaluation pull in opposite directions. Until an outside group publishes numbers, the public record is the vendor's clips plus journalists' impressions of the vendor's clips.

There is a second-order cost that rarely makes it into comparisons. A downloadable model is a dependency you own: 2.1 GB, a process to keep alive, no SLA, and a licence threshold to watch as you grow. A waitlisted model is not a dependency at all — it is a research result you can cite and cannot ship. For a team choosing a voice stack, those are not two points on one line. They are two different decisions.

What a voice agent needs that neither model supplies alone

Production voice systems are rarely one model. Whether the front end is duplex or request-and-response, the conversation keeps handing work to ordinary text models: a lookup, an extraction, a policy check, a reasoning call behind a held pause. That delegation layer is where cost variance lives, and where one provider's bad hour becomes your outage.

To be precise about our own coverage: LFM3-Audio-2B is not on OrcaRouter, and Dessa is not an API at all. What OrcaRouter serves is the text layer a voice agent delegates to — more than 200 models behind one key at provider list price with no markup added, automatic failover when a provider stumbles, a routing DSL that composes several models into one call, and model fusion when a single model's judgment is not enough. If you are prototyping with LFM3-Audio-2B locally, the tool-calling half of the loop is the part you can put behind that endpoint today; the audio half stays on your machine.

The test that would settle it

For LFM3-Audio-2B, the missing evidence is specific: an independent WER run on a standard set, a latency measurement on hardware that is not an M4 Max, and a reproduction of the UTMOS gain by someone who does not work for Liquid. The live-ASR follow-up in llama.cpp matters too, because batch transcription is the supported path today and English-only is the limit when streaming lands.

For Dessa, the list is shorter and harder: a held-out conversation set, end-to-end latency including the first audible reaction, interruption recovery, intelligibility under noise, a data and compute budget, and some measure of factual retention. Any one of those would turn a demo reel into an evaluation. All of them would make it a product.

Until then, the accurate description is narrow and useful. LFM3-Audio-2B is a 2B speech model you can download in a minute, that talks better than the model it replaces, listens about as well, and has no independent witnesses yet. Dessa is the most interesting voice demo of the month, with no way to check it. If you need to build this quarter, only one of those sentences ends in a deployment.