Discussion

Dessa versus LFM2.5: a closed speech demo meets weights you can actually use

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#8856

Block’s Dessa and Liquid AI’s LFM2.5-2.6B-Base represent very different bets on building speech products: one is a closed research preview, the other a downloadable model that teams can adapt. That difference may matter more to product builders than any benchmark score.

Watch Desk analysis

What happened

Dessa is described as a full-duplex speech model, trained on human conversation and reportedly fast enough to run on a MacBook Pro M4 Max. But it is not downloadable or available to measure: access is through a waitlist, with no published licence or price in the comparison. The account says Block announced it on 23 October.

LFM2.5-2.6B-Base, by contrast, is a 2.69-billion-parameter text model that the article says is available to download, can be fine-tuned, and uses Liquid AI’s LFM Open License. The article puts its runtime memory at about 2.5 GB and its context window at 131,072 tokens. It is a base checkpoint, not a ready-made conversational assistant: it needs further work to follow instructions or hold a dialogue. OrcaRouter’s comparison frames the choice as control and access versus a finished speech behaviour that users cannot yet inspect.

Why it matters

For a team building a product, access to weights means it can fine-tune, test and plan around a real artefact. A waitlist and unpublished terms offer less to evaluate, however appealing the demonstration. That makes these models difficult to compare as if they were interchangeable products.

There is a practical caveat on the LFM side: the article says commercial use above a $10 million annual-revenue threshold requires a separate licence. A known term can be checked with legal and procurement teams; a missing term is harder to plan around. And a text checkpoint is not a voice assistant in a small-model costume: speech recognition and synthesis would still be needed for an audio product.

Our read

This is less a verdict on which model sounds better than a useful contrast between a research preview and something builders can download and test. If reproducibility, customisation or deployment control is the priority, LFM2.5-2.6B-Base gives teams somewhere concrete to start. If the goal is natural, interruptible speech, Dessa’s reported behaviour may be compelling, but the public description does not let outsiders test that promise yet.

The comparison is published by OrcaRouter, which sells access to AI models. Its figures and descriptions should be read as the source’s account, not an independent head-to-head test. The useful takeaway is about access and product maturity, not a proven performance ranking.

What to watch

  • Whether Block makes Dessa available for independent testing, and on what terms.
  • Whether Liquid’s base checkpoint can be fine-tuned into a useful speech pipeline.
  • How commercial licensing affects teams considering LFM2.5-2.6B-Base.

Discussion spark: Would you build on weights you can inspect and adapt, or wait for a closed model that may deliver better speech behaviour?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.