Mercedes-Benz partners with Liquid AI to run non-transformer LFMs on-device in North American MBUX cars
Mercedes-Benz and Liquid AI announced a multi-year partnership (2026-04-23) to run embedded Liquid Foundation Models on-device — fast, private, no cloud dependence — in Mercedes models with third- and fourth-generation MBUX in North America, built on the MB.OS architecture. The release frames H2 2026 as a target: it "creates a path to a first production deployment for advanced speech technology," i.e. an initial speech deployment, not a confirmed full rollout. Liquid's LFMs are genuinely non-transformer (liquid neural networks / state-space models), making this one of the few production efforts off the attention architecture.
Why it matters (spine). Nearly every shipping foundation model runs on the transformer's attention mechanism; Liquid AI's LFMs do not — they are liquid neural networks and state-space / linear-complexity operators, a family cheap enough to run on the car's own silicon rather than a datacenter. A named automaker committing multi-year to embed such models on-device — fast, private, no cloud round-trip — is a concrete data point that value can accrue off the frontier: at the edge, on hardware the user already owns, and on a non-attention architecture. For the machine-layer thesis it shows a defensible position being built where most of the field is not looking.
What would prove it wrong (falsifier). Production slips indefinitely, or independent testing shows the on-device model underperforms the cloud badly enough to force a fallback to cloud reliance — collapsing the "independent AI without dependence on the cloud" claim.
Live-status check + triangulation. Quoted from the joint Liquid AI / Mercedes-Benz press release dated 23 April 2026: "Embedded Liquid Foundations Models (LFM) deliver fast, private and independent AI without dependence on the cloud." Fetched from the liquid.ai primary in-session. Corroborated by Mercedes-Benz USA's newsroom (media.mbusa.com), the BusinessWire wire release (20260423009970), and Automotive World's trade coverage; Liquid AI's own LFM architecture post supplies the non-transformer detail. Note: the Mercedes-Benz Group collaboration landing page originally cited is a lighter first-party summary and returned HTTP 403 in-session, so the canonical dated release is taken from liquid.ai / media.mbusa.com / BusinessWire.
Correction carried forward (discipline).
- The draft title said the model is "shipping on the car itself"; the primary says the partnership "creates a path to a first production deployment for advanced speech technology, targeting the second half of 2026" — a stated target, not yet shipping.
- The draft capsule said "targeting production in H2 2026" broadly; the primary scopes H2 2026 to a FIRST production deployment for advanced SPEECH technology, not the full model or a whole-fleet rollout.
- "Post-transformer" / "off the transformer" is accurate but is NOT stated in the Mercedes/Liquid press release itself; it comes from Liquid AI's own architecture (liquid neural networks + state-space / linear-complexity operators) and is triangulated, not sourced from this primary.
- The originally cited URL is a Mercedes-Benz Group collaboration landing page (first-party but a lighter summary) and returned HTTP 403; the canonical dated joint release lives at liquid.ai, media.mbusa.com, and BusinessWire.
- The draft's claimed date, 2026-04-23, was CONFIRMED correct against the primary release date (no change).
Card the shift: private, on-device, fine-tuned foundation models are shipping to consumer hardware. Flag — don't repeat — Hasani's unverified specifics (the sub-1GB / $60-chip numbers aren't in the official release).
domains: machine-layer · on-device-ai · non-transformer · automotive · foundation-models · edge-inference