The artefact size Ramin Hasani named on the Latent Space podcast is not primarily a model specification — it is a channel selector, and the channel it selects already belongs to the customer.
LIQUID AI × MERCEDES-BENZ — KEY FACTS
600 MB
Model artefact size named by Hasani on Latent Space
H2 2026
Targeted initial production deployment — planned, not shipped
302
Neurons in C. elegans — the design constraint Hasani cites as architectural origin
23 Apr 2026
Joint announcement date, multi-year partnership, Gen 3 & 4 MBUX, North America
On the Latent Space episode titled “A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI,” Hasani described the artefact his company intends to place inside a vehicle: “this model itself is about 600 megabyte so imagine like if you have like a 600 megabyte intelligence that goes inside every car you can do an over-the-air update or OTA — you can do over-the-air update of every car on the planet.” That sentence deserves to be read slowly, because it does two things at once: it names a size, and it immediately names the distribution mechanism that size makes available.
The commercial context for that statement is a joint announcement made on 23 April 2026 by Liquid AI and Mercedes-Benz, naming Jörg Burzer — Member of the Board of Management of Mercedes-Benz Group AG and Chief Technology Officer for Development and Procurement — and Ramin Hasani, Co-Founder and CEO at Liquid AI. The announcement describes a multi-year partnership to scale embedded in-car intelligence for third- and fourth-generation MBUX in North America, with Liquid Foundation Models processing speech, language understanding and reasoning locally — characterised as fast, private and independent AI without dependence on the cloud — and a path toward initial production deployment targeted for the second half of 2026. That deployment is targeted and planned, not shipped. The system is not in cars, has not launched, and is not on the road.
On the programme, Hasani described the first deployment as going to happen this year and covering North America Mercedes-Benz cars that are generation 3. The joint announcement refers to third- and fourth-generation MBUX in North America, with a targeted H2 2026 path toward initial production. Both accounts are reported here as they were made; neither is adjudicated against the other, and neither is described as wrong. Liquid AI has published LFM2 post-trained checkpoints at 350M, 700M, 1.2B and 2.6B parameters. A parameter count is not a file size, and no correspondence between any of those figures and the 600 megabytes is established here.
The key insight: Six hundred megabytes is not primarily a model specification. It is a channel selector. The size of the artefact determines which delivery infrastructure can carry it — and a model that fits inside an existing over-the-air update rides a channel the customer already owns, already operates, and already uses for other purposes. That is a different business from one that requires the vendor to build, power, and defend the channel itself.
Size decides which channel you can use. That is a distribution decision before it is an engineering one.
The Structural Read
The Business Engineer lens here is size as a distribution decision. Notice the structure of Hasani’s sentence: it moves in one step from the size of the artefact to the mechanism by which it reaches every unit in a fleet. That is not a coincidence of phrasing. The two are the same thought.
A model that runs in a data centre requires a channel the vendor has to build, power and defend — physical infrastructure, interconnection, capital expenditure, and the permission of whoever governs each site. A model that fits inside an existing over-the-air update rides a channel the customer already owns, already operates, and already uses for other purposes: firmware patches, map updates, feature rollouts. Those are different businesses with different capital requirements and different people who have to say yes before the model reaches the end-user. This is a general property of artefact size and delivery channel. It is not a claim about Liquid AI’s costs, margins, strategy or prospects — none of which is established here — and it is not a claim that running on the device is better, cheaper, faster or more likely to succeed than any alternative approach.
The architectural origin story reinforces this reading. On the programme, Hasani said of the nematode C. elegans: “With 302 neurons it could control 95 muscle cells better than any robotic systems that we had on the planet.” That is his characterisation, dated by him to 2016 and 2017. It is not presented here as an established scientific or engineering finding, and no biological or robotics claim is made. What matters for the structural read is what the characterisation reveals about the design constraint: a system selected for behaving well with very few units is a system optimised for a hard parameter budget, and a hard parameter budget is precisely the condition imposed by anything that has to fit inside a vehicle and propagate across a fleet via an existing update channel. The lineage is not decoration. It names the constraint the architecture was chosen to satisfy. That is an observation about what the architecture was selected for — not a claim that the approach works, is superior, or will succeed.
Business Engineer Framework — The Channel the Customer Already Owns
Size determines which gate you have to pass through
A data-centre model requires the vendor to negotiate access to infrastructure before the model reaches any user. A model that fits the OTA envelope skips that gate entirely — the customer’s existing update pipeline is the distributionnetwork, and it is already funded, already live, and already trusted by the regulator who approved it. Different artefact size, different permission structure, different capital stack. This is a structural observation about how size and channel interact; it is not a claim about any specific company’s costs, strategy, or likelihood of success.
The third signal worth recording is the evaluation argument. On the programme, Hasani said: “your evals are going to become obsolete. Static evals are going to become obsolete. So you need to have some sort of a dynamism like there and this would only happen if you have a continuously evolving system.” That is his argument; it is neither endorsed nor refuted here.
What makes it worth recording precisely is that it is the third independent version of the same structural claim to surface this week. Separately and independently, two companies funded this week sell, respectively, the reading of production agent trajectories to catch failures introduced by a model upgrade, and continuous validation of controls in place of a point-in-time audit. No connection, causation or coordination is claimed between any of these parties, and convergence is not evidence of coordination. The structural observation stands on its own: when the thing being measured keeps moving, a measurement taken once describes a state that has already been left behind. That is true whether the moving thing is a model, a permission set, or a user population.
Three Implications
IMPLICATION 1 — THE DEPLOYMENT IS STILL TARGETED
The joint announcement of 23 April 2026 describes a path toward initial production deployment targeted for H2 2026. Hasani’s on-programme account describes the first deployment as happening this year across North America generation-3 cars. Both statements are reported as made. The system is not in cars, has not launched, and is not on the road. Keeping that distinction visible is not a caveat — it is the only honest way to read a targeted deployment timeline, and the distinction matters for anyone reasoning about what has been demonstrated versus what has been announced.
IMPLICATION 2 — ARTEFACT SIZE RESHAPES WHO MUST SAY YES
A model delivered via OTA inherits the approval structure of the update mechanism it rides — firmware governance, homologation, the OEM’s own security review. A model delivered via a vendor-operated cloud requires a separate approval chain: data-residency, latency SLA, uptime contract, and the permission of whoever governs the infrastructure in each market. These are different sets of people with different incentives and different timelines. Artefact size, by determining which channel is available, also determines which approval chain applies. That is a structural property of the choice, not a prediction about any outcome.
IMPLICATION 3 — THE SNAPSHOT IS THE WRONG INSTRUMENT FOR A MOVING TARGET
Hasani’s argument that static evaluation suites will become obsolete as systems evolve continuously is the third independent articulation this week of a single structural claim: a point-in-time measurement cannot track a system that is moving. If the model updates over the air, the evaluation must also travel with it. This is not a product recommendation. It is an observation about the shape of any production system where the artefact changes after deployment — the measurement instrument has to be redesigned alongside the delivery mechanism, or it stops describing the thing it claims to measure.
Business Engineer Framework
The Map of AI — Where the Artefact Sits Determines What the Business Looks Like
The Map of AI traces how value pools form across nine layers of the AI stack — from foundation model to distribution channel to the end-user context. The Liquid AI / Mercedes-Benz story is legible through that map: a model compressed to fit the OTA envelope is not just an engineering decision, it is a choice about which layer of the stack the company occupies and which capital requirements follow. Understanding where an AI artefact sits in the stack — and why its size places it there — is the first move in any structural analysis of the AI industry.
Six hundred megabytes is a business model before it is a model specification: the size of the artefact selects the delivery channel, the delivery channel determines whose infrastructure carries the risk, and the infrastructure question determines which people must approve the deployment before a single car is updated. The Mercedes partnership is targeted for H2 2026 — planned, not shipped, and reported here only as such — but the structural logic Hasani described on Latent Space is already live: a model small enough to ride an existing OTA pipeline inherits the permissions, capital, and governance of a channel the customer built and owns, and that is a categorically different commercial position from one that requires the vendor to construct and defend its own path to the edge.
This article is businessanalysis only. It is not investment advice, no view is expressed on any security, and no recommendation of any kind is made.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
The Mercedes-Benz deployment described above is targeted and planned, not shipped. Nothing above should be read as saying the system is in cars, has launched, is on the road, or is available to drivers. Ramin Hasani’s statements about timing and scope were made on the programme and are reported as he made them, alongside the scope and timing set out in the joint announcement of 23 April 2026. Neither account is adjudicated against the other here and neither is described as wrong. A parameter count is not a file size. The roughly 600 megabytes Hasani describes is a file size; Liquid AI’s published LFM2 checkpoint sizes are parameter counts. No correspondence between the two is established here. The comparison between a nematode’s 302 neurons and robotic systems of 2016 and 2017 is Hasani’s characterisation, not an established scientific or engineering finding, and no biological or robotics claim is made above. Where other companies are noted as having made similar arguments about static measurement, no connection, causation or coordination is claimed; convergence is not evidence of coordination. Nothing above claims that running models on the device is better, cheaper, faster or more likely to succeed than any other approach, names or compares any competitor, or makes any claim about any company’s costs, margins, strategy or prospects. No benchmark, accuracy, latency, throughput, vehicle count, volume, contract value, price, funding, valuation, headcount or compute figure appears above, no outcome or customer effect is claimed, and nothing is predicted. This is businessanalysis. It is not investment advice, no view is expressed on any security, and no recommendation is made.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
