Dessa vs LFM2.5 2.6B Base: A Checkpoint You Can Own vs a Voice You Can Only Hear

LFM2.5-2.6B-Base is a 2.69-billion-parameter file that fits in about 2.5 GB, runs on a phone, and comes with a licence that lets you fine-tune it and ship the result. Dessa is a full-duplex speech model that Block announced on 23 October 2026, trained only on human conversation, and reportedly fast enough to run on a MacBook Pro M4 Max. You cannot download Dessa, fine-tune it, measure it, or call it. It is a waitlist.

Putting these two in the same sentence is unfair in an obvious way: one is a pretrained text checkpoint, the other is a finished conversational behaviour. But the unfairness is the useful part. The question a team actually faces is not "which model is better" — it is "which model can I build a business on," and on that question the gap between a downloadable base checkpoint and a closed research preview is larger than any benchmark difference.

What LFM2.5-2.6B-Base actually is

Liquid AI released the LFM2.5 family on 4 August 2026. The Base suffix matters more than the parameter count: this is the pretrained checkpoint, not the instruction-tuned assistant. It has read 34 trillion tokens and learned the statistical structure of 16 languages, but it has not been taught to follow instructions, hold a dialogue, or use tools. If you prompt it cold, you get completion, not conversation.

The architecture is a hybrid. Thirty layers, of which 22 are double-gated short-convolution blocks and eight are grouped-query attention, a design Liquid has been refining since the first LFM2 generation. The vocabulary is 128,000 tokens, doubled from the previous generation through an in-place tokenizer extension aimed at non-Latin scripts. Context is 131,072 tokens, reached through a dedicated mid-training extension rather than at pretraining time.

The instruction-tuned sibling, LFM2.5-2.6B, is the one with the published benchmark table: AIME25 51.87, LiveCodeBench v6 59.41, IFBench 59.17, BFCLv4 56.88, ToolSandbox 77.83, Claw-Eval English 62.85. Those are vendor-reported and they belong to the post-trained model, not the base. Liquid itself advises against using the family for agentic coding or knowledge-heavy work, which is a useful honesty signal in a market where every small model is described as punching above its weight.

Why comparing a base checkpoint to a finished behaviour is still worth doing

A base model is a starting point. Its value is what you can make from it: continued pretraining on your domain, a supervised fine-tune on your transcripts, a distilled student for a narrow task. Dessa is the opposite kind of artifact. It arrives as a behaviour — turn-taking, backchannel, emotional timing — and none of it is exposed for modification.

So the comparison is really about control. LFM2.5-2.6B-Base gives you the weights, the licence, and the obligation to do the work. Dessa gives you a demo and the possibility of access. For a research lab, the second can be more valuable. For a product team, the first is the only one that produces a deliverable.

The spec sheet, with both sides on one line

LFM figures are from Liquid's model card and announcement; Dessa figures are from Block's page. All vendor-reported numbers are labelled as such.

• Parameters — LFM2.5-2.6B-Base 2.69B total vs Dessa undisclosed.

• Architecture — LFM2.5-2.6B-Base a hybrid of short-convolution and grouped-query attention, 30 layers vs Dessa end-to-end speech-to-speech, architecture unpublished.

• Context — LFM2.5-2.6B-Base 131,072 tokens vs Dessa not stated.

• Modality — LFM2.5-2.6B-Base text in, text out vs Dessa audio in, audio out.

• Weights — LFM2.5-2.6B-Base downloadable, including GGUF, ONNX, and MLX builds vs Dessa none.

• Licence — LFM2.5-2.6B-Base the LFM Open License v1.0, commercial use with conditions vs Dessa no licence published.

• Memory — LFM2.5-2.6B-Base under about 2.5 GB at runtime vs Dessa a MacBook Pro M4 Max or better, per Block.

• Speed — LFM2.5-2.6B-Base about 220 tokens per second on an Apple M5 Max, about 30 on phones, vendor-reported vs Dessa "faster than real time," with no figure.

• Price — LFM2.5-2.6B-Base free weights vs Dessa no price.

• Availability — LFM2.5-2.6B-Base generally available since 4 August 2026 vs Dessa a waitlist.

The device budget is the whole argument

Liquid's pitch is deployment economics: a model small enough for a Raspberry Pi, fast enough for a phone, cheap enough to run without a GPU bill. The claimed throughput — about 220 tokens per second on an M5 Max, 113 on a Ryzen AI Max+ 395, and roughly 30 on a phone — is the kind of number that determines whether an on-device assistant feels instant or broken. Thirty tokens per second is faster than reading speed, which is the relevant threshold for a chat interface.

Dessa's deployment claim is a laptop costing well over $2,000 with a top-end Apple chip. That is still local, and local has real advantages: no per-token bill, no network round trip, no data leaving the device. But it is a different market. A 2.6-billion-parameter model can ship inside an app; a model that needs an M4 Max is a developer tool or a desktop feature. If your product runs on phones, this comparison ends at the memory line.

Licensing is the real divide

The LFM Open License v1.0 is not an OSI-approved open-source licence, and it is worth saying so plainly. Commercial use is permitted, but organisations above a revenue threshold — Liquid's terms put it at $10 million in annual revenue — need a separate commercial licence. That is a real constraint. It is also a known constraint, with a document, a threshold, and someone to negotiate with.

Dessa has no licence because Dessa has no distribution. You cannot comply with terms that do not exist, and you cannot build a procurement case around a waitlist. For an enterprise, "custom licence with a revenue threshold" is a legal review. "No published terms" is a blocker.

If you want a voice agent on a phone today

Neither model is a complete answer, but the LFM side is at least a starting point. A phone voice agent needs three pieces: speech recognition, a language model, and speech synthesis. LFM2.5-2.6B-Base can be the middle piece if you fine-tune it — the base checkpoint is a text model, so it will not hear anything on its own, and its 131K context is enough to hold a long conversation history plus retrieved notes.

The fine-tune is the interesting option. With weights in hand, you can train on your own domain conversations and evaluate the result on your own terms. That is precisely the experiment Dessa's announcement invites and does not permit: take a small speech-capable stack, train it on real conversations, and see whether emotional timing emerges. Block says it did this from scratch on one GB200 NVL72 rack. A team with a few GPUs can run the cheap version of the question on an LFM base and learn something in a week.

If you would rather not run the language model on-device, OrcaRouter puts one API in front of 200+ models with 0% markup and automatic failover, so the reasoning step can live in the cloud while audio stays local. To be clear about scope: LFM2.5 2.6B Base is not in our catalogue — it is a download, and that is the point of it. We route hosted models; this one is meant to run where you put it.

Who should pick which

• You need to own the model, the weights, and the roadmap — LFM2.5-2.6B-Base, then fine-tune.

• You need offline operation on cheap hardware — LFM2.5-2.6B-Base.

• You need a text or agentic core under a few gigabytes — the post-trained LFM2.5-2.6B, not the base.

• You need live, interruptible, emotionally expressive speech — Dessa is the only one of the two aimed at it, and it is not available.

• You need a licence you can sign this quarter — LFM2.5-2.6B-Base, with the revenue clause checked.

• You need to publish a reproducible result — LFM2.5-2.6B-Base, because reproducibility requires a file.

The experiment that would settle it

Take 1,000 hours of conversational audio with transcripts, fine-tune LFM2.5-2.6B-Base on the text side, and bolt on a strong speech recogniser and synthesiser. Measure interruption recovery and emotional appropriateness against a turn-based baseline. If the result approaches what Block's clips suggest, then Dessa's advantage is narrower than the demo reel implies: it may be an architecture win, but not an unassailable one.

If it does not, the case for speech-native training gets stronger, and the absence of a Dessa weight file becomes the most important fact in this comparison. Either way, the test is runnable on the open side and not on the closed one. That asymmetry is the answer to the headline question, and it will stay the answer until Block ships something a customer can hold.