LFM3-Audio-2B vs Sesame Preview: A Download and an App Are Not the Same Choice

Sesame Preview is the voice people describe as human. LFM3-Audio-2B is a 2B checkpoint you can download and run on a laptop. Put them in the same sentence and something useful happens: the comparison stops being about which one sounds better and becomes two separate questions — can you ship it, and can you measure it? Sesame Preview fails the first. LFM3-Audio-2B, which Liquid AI shipped on 27 October 2026 inside the six-model LFM3 family, fails the second. Neither failure is a scandal. They are different kinds of product, and most teams discover which one they need only after they have started building.

The asymmetry is worth stating plainly before the details. Sesame's production voice has no API, no model ID and no enterprise tier; it lives inside an iOS and Android app with four named agents, memory, connectors and, since October, the ability to orchestrate other agents. LFM3-Audio-2B has no app at all. It has weights, a model card, a merged llama.cpp pull request and a 2.1 GB Ollama build. One is a place to spend an afternoon. The other is a file you can put in a product.

Question one: can you ship it?

For LFM3-Audio-2B the answer is yes, with caveats. The LFM Open License v1.0 permits commercial use below $10M annual revenue and requires a paid licence above it. Weights are on Hugging Face; Transformers 5.20.2 or newer with trust_remote_code is the Python path; llama.cpp merged support in PR #31257; MLX 4-bit covers Apple silicon; ONNX Runtime covers portability. Input is text or 16 kHz audio, output is text, 24 kHz speech, or interleaved text and speech. Context is 32K tokens covering both directions — not the 256K figure that belongs to the text model — and the model lists 18 languages across 19 locales, with live ASR English-only.

For Sesame Preview the answer is no, and the reason is structural rather than commercial. There is no endpoint for the production voice. There is no price, because there is nothing to buy. The app is free, calls run to a 30-minute cap that drops to five minutes when logged out, and the four agents are the product. If you want Sesame-shaped behaviour in your own application, you cannot have it.

What Sesame has released is CSM-1B, an Apache-2.0 research checkpoint. It is a synthesis block, not an assistant: a Llama backbone predicts the first Mimi codebook and a smaller decoder predicts the rest, rendered through the Mimi codec at 24 kHz. It is English-only, clones a voice from a 10-to-15-second reference clip, and cannot generate text — you pair it with a language model. Its Hugging Face card says the quiet part directly: a fine-tuned variant of CSM powers the interactive demo, and that variant is not the checkpoint you can download. The card has roughly 130,000 downloads in the last month and about 2,500 likes, which tells you people are experimenting with it. It does not tell you they are getting the voice from the app.

Question two: can you measure it?

LFM3-Audio-2B publishes a six-row vendor table, all against LFM2.5-Audio-1.5B, with no named test set or harness. The rows that matter here are English TTS UTMOSv2 at 3.83 against 3.22, VoiceBench at 73.13 against 55.17, audio understanding MMAR at 50.60 against 25.85, English ASR WER at 6.23 against 6.13, multilingual WER/CER at 6.03, and an internal tool-calling eval at 58.50. The UTMOS move is large — 0.6 points on a scale where 4.0 is usually treated as near-human — and the ASR row is flat, which is the most credible thing in the table: a vendor that reports no improvement on one axis is at least not optimising the table for a story.

None of it has an independent witness. We could find no third-party reproduction of any row, and the numbers compare a model with its own predecessor, not with the field. Against the other entries in Liquid's own table, LFM3-Audio-2B is a good small model rather than a leader: Voxtral Mini 3B posts 5.51 WER and NVIDIA Audex 2B posts 5.99, Qwen2.5-Omni 3B takes audio understanding at 54.41, and Gemma 4 E4B takes tool calling at 72.40.

Sesame Preview has no comparable number at all. Its quality claim is a listening impression — one strong enough that it became the reference point for conversational voice in 2026, and one that a third party has never scored for the production model. That is not evidence the voice is bad. It is evidence that the thing everyone praises is the one thing in this matchup nobody can measure.

This inversion is the rare part. Normally the closed product is the one with the track record and the open checkpoint is the one nobody has evaluated. Here the open model is the only one you can evaluate, and the closed product is the only one with a real user base. Both statements are true at once, and neither settles whether the voice is good.

Two codecs, two design bets

The most technical difference between these systems is also the most interesting, because both companies bet on a codec and the bets point in opposite directions.

LiquidCodec packs 80 ms of speech into 12 discrete tokens and decodes back to 24 kHz with streaming support. Liquid measures it at about 87× real time on a single CPU thread of an M4 Max, and the causal audio encoder is what allows prefill while a person is still speaking — the source of the "as low as 50 ms" time-to-first-token claim. Those two figures are vendor measurements on named Apple hardware with no harness and no percentile. They describe a best case. A phone, a Raspberry Pi or a shared x86 VM will not hit them, and nothing in the release says what it will hit instead.

Sesame's CSM uses Kyutai's Mimi codec at 24 kHz, with a Llama backbone predicting the first codebook and a smaller depth decoder predicting the rest. The design goal is expressive synthesis from a short reference clip, and the app's reputation suggests it works. The gap between CSM-1B and the production voice is the whole argument: the architecture is public, the fine-tune that makes it sound like Sesame is not.

What that means for a builder is simple. Liquid gives you the whole pipeline and a licence; you supply the evaluation. Sesame gives you a research artifact and a demo; you supply the product, and you cannot put the demo inside it.

Where the two actually overlap

There is exactly one overlap, and it is voice output. Both can speak. The differences inside that overlap are practical:

• Languages — LFM3-Audio-2B claims 18 languages across 19 locales, with translation strongest to or from English. Sesame Preview is English-only officially; the app's own guidance lists French, Italian, German, Spanish, Chinese, Japanese and Korean as "early". CSM-1B is English-only.

• Voice control — LFM3-Audio-2B: prompt-selected voice, no published cloning workflow in the card. CSM-1B: cloning from a 10-to-15-second reference clip, which is its main feature.

• Conversation — LFM3-Audio-2B: request-and-response with streaming; no barge-in claim. Sesame Preview: built around hesitation, interruption and turn-taking, which is the quality that made its reputation.

• What you can run — LFM3-Audio-2B: 2.1 GB, CPU-capable, offline. Sesame: an app, online, with a call cap.

If your product needs a voice that can be interrupted, this is not a close comparison: Sesame has the behaviour and Liquid does not claim it. If your product needs a voice in Swedish or Dutch, or needs to run without a network, Sesame cannot help and LFM3-Audio-2B can.

What a router can and cannot do here

Neither model is on OrcaRouter, and it is worth being exact about why the usual routing pitch does not reach the speech layer. LFM3-Audio-2B is a checkpoint you run yourself; Sesame Preview is an app. There is no hosted audio endpoint of either to fail over between.

The layer above them is where routing earns its keep. A voice product built on either model still hands work to text models — summarising a call, extracting a booking, deciding the next question — and that delegation is where cost variance and provider outages live. OrcaRouter serves more than 200 text and multimodal models behind one key at provider list price with no markup added, with automatic failover and a routing DSL that can compose several models into one call. It also lets you run model fusion when a single model's judgment is not enough. For a team prototyping a voice loop, that means the text half can be swapped and compared without touching the audio integration.

The decision

If you are calibrating taste — deciding how much warmth your product should have, how long a pause is comfortable, how quickly an agent should yield the floor — Sesame Preview is free, genuinely excellent, and the right place to spend an afternoon. Use it as a reference. Do not plan an integration around it, because there is nothing to integrate with.

If you are shipping, LFM3-Audio-2B is a real candidate with real caveats: vendor-only numbers, English-only live ASR, a 32K context, and a licence threshold to watch. Test it on your own recordings before you believe the UTMOS gain, and test speech output in the languages you actually need rather than trusting the locale count.

What would change this picture is specific. A Sesame API for the production voice would make it the strongest candidate in the category overnight. An independent UTMOS or preference test for LFM3-Audio-2B would tell you whether 3.83 is a real step toward natural speech or a number measured against a weak predecessor. Until one of those lands, the correct summary is that one of these is a product you can admire and the other is a model you can own — and only one of those is a decision you can act on this week.