Liquid AI’s 600MB Mercedes model turns file size into a distribution strategy

Ramin Hasani has put a number on the intelligence Mercedes-Benz plans to build into its cars: about 600 megabytes, small enough to reach the whole fleet as an ordinary over-the-air update. The figure came up on the Latent Space podcast, and its significance is less benchmark than business model. A model that fits the update channel a carmaker already runs inherits that channel's reach, approvals and costs, which is a different commercial position from one that needs its own pipe to the vehicle. Nothing is on the road yet. The partnership was announced jointly on 23 April 2026, with initial production deployment targeted for the second half of this year in North America. This is a plan with a date, not a shipped system.

Watch Desk analysis

What happened

Hasani, Liquid AI's co-founder and chief executive, described the artefact plainly: the model is about 600 megabytes, and that size means you can push an over-the-air update to every car on the planet. The joint announcement he fronts alongside Jörg Burzer, Mercedes-Benz's board member for development and procurement, describes a multi-year partnership running Liquid Foundation Models locally in third- and fourth-generation MBUX, handling speech, language understanding and reasoning without the cloud.

On the programme, Hasani said the first deployment lands this year on generation 3 cars in North America. The announcement covers generations 3 and 4. Both accounts stand as made, and the gap between them is worth watching rather than adjudicating.

One distinction deserves carrying. Liquid AI's published LFM2 checkpoints are measured in parameters, from 350 million to 2.6 billion, and a parameter count is not a file size. The 600 megabytes is the chief executive's figure for the artefact, and the two scales have not been publicly reconciled.

The origin story, per Hasani, is a worm. He credits the nematode C. elegans with controlling 95 muscle cells from just 302 neurons, his own characterisation rather than settled science, but it names the design constraint: a system selected to behave well on a very tight budget, which is precisely the condition imposed by anything that must live inside a car.

Why it matters

Artefact size decides which door a model can walk through. A model that needs a data centre requires its vendor to build, power and defend a channel to the vehicle, then win permission from whoever governs each market. A model that fits inside an existing over-the-air update rides infrastructure the customer already owns, operates and has already had approved. Different size, different permission structure, different capital stack.

There is a second consequence for anyone measuring these systems. Hasani argues static evaluation suites will go obsolete for models that keep updating themselves; if the artefact changes over the air, a one-time score describes a car that no longer exists. Evaluation would have to travel with the update, or stop describing the thing it claims to measure.

Our read

The distribution logic is sound and unusually concrete: 600 megabytes is a channel selector before it is a model specification. Treat the number as the claimant's own, though. It comes from the chief executive of the company selling the model, the deployment is targeted rather than shipped, and the analysis framing this story is business reading rather than independent measurement. The sensible stance is interest with a calendar entry. If the second half of 2026 closes with 600MB models updating Mercedes fleets, edge AI will have found its distribution template. If it slips, the gap between announcement and road becomes the story.

What to watch

• The H2 2026 target
  Whether the first generation 3 deployment in North America actually ships this half.
• The 600MB figure
  Whether the artefact size is confirmed independently, and which hardware it runs on.
• Generation 4 scope
  Whether fourth-generation MBUX joins the first rollout or follows later.
• Evals on the move
  Whether Liquid AI publishes how it tests a model that updates itself over the air.

Discussion spark: For edge AI, is the winning artefact the biggest model the hardware allows, or the biggest that still fits the customer's update pipeline?

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.