AI AI Toolkit
AI Newsai-models

Liquid AI 发布 LFM3 多模态开源权重模型家族,2B 级覆盖文本、视觉、音频、omni 和决策模型

Liquid AI 模型与工程博客(网页)2026-10-26T16:00:00.000Z

Core Highlights

Liquid AI has released LFM3, a new family of multimodal models built around the theme of "specialized intelligence" and on-device deployment. The company, a spinoff from MIT, continues its line of liquid neural networks (LNNs) — but the strategic goal here is not to fight the general-purpose giants on parameter count. Instead, LFM3 is designed to be small, fast, and capable of running directly on phones and in-vehicle computers. In plain terms, Liquid AI is betting that a huge slice of real-world AI value will come from models that live on the device rather than in a distant data center, where latency, cost, and privacy all work against the user.

What the Release Brings

LFM3 handles multimodal inputs such as text and images, and it emphasizes low-latency inference on edge devices — phones, cars, and edge chips. The contrast with cloud-centric AI is sharp: instead of shipping every query to a server farm, an on-device model can keep working with no network connection, under tight power budgets, and with strong privacy guarantees. That makes it a natural fit for real-time perception and local decision-making, where a round-trip to the cloud is either too slow, too expensive, or simply impossible in the field. For a factory floor or a moving vehicle, "always online" is not a safe assumption, and that is exactly the gap LFM3 targets.

Technical Details

Liquid AI's core technology traces back to liquid neural networks and state-based dynamical-systems modeling. The models have a continuous-time modeling ability over input sequences, which keeps parameter count and compute demand relatively contained while preserving expressive power. With LFM3, the team extends this thinking to multimodal data, so an edge device can understand an image and reason about it in tandem with text — without calling home. That is a meaningful architectural choice: it treats efficiency and capability as a single design problem rather than a tradeoff resolved by brute-force scaling, and it is a direct continuation of research the lab has published for years.

How It Compares

Against cloud behemoths like GPT and Claude, LFM3's differentiation is "small but specialized, and actually runnable." It is not trying to win a benchmark brawl; it is trying to win the places where a 700-billion-parameter model simply cannot go. Its positioning is similar to the neural-engine ecosystems pushed by Qualcomm and Apple — both target the blue ocean of edge AI, pushing intelligence down into the device instead of concentrating it in the data center. The difference is that Liquid AI sells the models and the know-how, not the silicon, which lets it partner across hardware vendors rather than competing with them.

Industry Impact and Use Cases

The value of on-device multimodal models is obvious for privacy, latency, and offline reliability. Concrete uses include in-car assistants that react to the cabin without sending audio to the cloud, industrial inspection cameras that flag defects locally, and wearables that process health signals on the wrist. As smartphone NPUs grow more capable, these "specialized small models" may become the invisible workhorses of AI adoption — the quiet majority running beneath the flagship models that grab the headlines. For developers, LFM3 is a clear signal that the edge is no longer a stripped-down afterthought but a first-class deployment target, and that building for constrained hardware is becoming a distinct, valuable skill set in its own right. The broader takeaway is that the future of AI compute is not one giant cloud but a federation of small, specialized models living close to where data is born.

Practical Takeaways

For product teams, the signal from LFM3 is that "small and on-device" is no longer a compromise but a design target with its own advantages: no per-call API bill, no data leaving the device, and no dependency on a network that may not exist at the point of use. The practical move is to audit which of your AI features actually need a giant cloud model and which could run locally on a phone NPU or an edge box, then prototype the latter to cut both cost and latency. Developers should also start learning the constraints of constrained hardware — quantization, pruning, and power budgets — because that skill set is becoming as valuable as knowing how to prompt a frontier model. For investors and strategists, Liquid AI's path is a reminder that the edge is a contested but open lane where a focused startup can still differentiate against the hyperscalers. The risk, of course, is that Apple and Qualcomm eventually absorb this category into their silicon and tooling; Liquid's defense is staying model-agnostic and partner-friendly rather than locked to one chip.

Wider Context

The bigger story is that AI compute is splitting into two economies: a small number of enormous cloud models for the hardest problems, and a widening fleet of specialized on-device models for everything else. Liquid AI is betting on the second economy, and the early enterprise interest suggests the bet is not crazy.