The Rise of Liquid Nanos: Rethinking Agentic AI

Posted on September 26, 2025 at 11:25 PM

The Rise of Liquid Nanos: Rethinking Agentic AI

What if less is the future of powerful AI?

Imagine your phone, laptop, car, and smart home each running dozens of tiny AIs — all zeroing in on specific tasks — instead of shoving every request into a giant cloud-model. That might sound counterintuitive in an age obsessed with ever-bigger foundation models. Yet MIT spin-off Liquid AI argues exactly that: we may have gotten “agentic AI” wrong all along by overemphasizing monolithic models. Their answer? Liquid Nanos — compact, task-specific models built for the edge.


From “one model to rule them all” → “many tiny agents”

In many current agentic AI systems, the pattern is:

  1. Take a generalist foundation model (e.g. GPT, Claude, Gemini)
  2. Narrow its behavior via prompting, memory, caching, fine-tuning
  3. Serve multiple tasks via the same large model

That works — until latency, cost, privacy, connectivity, or customization become constraints. Liquid AI proposes flipping that paradigm: ship intelligence to devices rather than shipping data to remote clouds. ([Venturebeat][1])

Liquid’s “Nanos” are models ranging from ~350 million to 2.6 billion parameters, each specialized for a task (e.g. data extraction, translation, tool invocation, math). ([Venturebeat][1]) Because they’re lightweight, they can run on field devices, laptops, even sensors and robots, without needing constant connectivity. ([Venturebeat][1])

What kinds of Nanos are already available?

Liquid AI rolled out six task-oriented models in its Liquid Nanos lineup:

  • LFM2-Extract (350M / 1.2B): structured data extraction from unstructured text
  • LFM2-350M-ENJP-MT: English ↔ Japanese translation
  • LFM2-1.2B-RAG: optimized for retrieval-augmented question answering
  • LFM2-1.2B-Tool: precision in tool / function calling
  • LFM2-350M-Math: math reasoning with controlled verbosity
  • Luth-LFM2 series: community fine-tunes (e.g. French) while preserving English capabilities ([Venturebeat][1])

Despite their small sizes, some Nanos outperform much larger models in benchmarks of accuracy, faithfulness, and syntactic validity. For example, the extraction model outperformed the much larger “Gemma 3 27B” in structured output tests. ([Venturebeat][1])

The licensing, deployment & trade-offs

  • Liquid Nanos can be downloaded and deployed via the Liquid Edge AI Platform (LEAP) (iOS, Android, laptops). ([Venturebeat][1])
  • They’re also available on Hugging Face under a custom LFM Open License v1.0. ([Venturebeat][1])
  • The license is free for individuals, researchers, nonprofits, and companies with < $10M revenue (with attribution and documentation requirements). Larger enterprises must negotiate separate commercial terms. ([Venturebeat][1])

Liquid AI’s philosophy is that many lightweight, specialized models lead to better latency, better privacy, lower cost, and more flexibility — especially in domains with connectivity or resource constraints. ([Venturebeat][1])


Why Liquid Nanos Matter—and Why the AI World Should Listen

  1. Latency & connectivity resilience Large, cloud-hosted AI can lag when networks are slow or unavailable. Tiny on-device models bypass that.
  2. Privacy & data control Sensitive data doesn’t have to leave the device.
  3. Cost efficiency Running inference on small models is much cheaper long-term than always hitting a large model in the cloud.
  4. Modularity & specialization Instead of one “jack-of-all-trades” model, you get multiple masters of small tasks.
  5. Scalability across devices Whether it’s phones, edge sensors, or robotics, these models can scale into constrained environments.

In short: while the AI industry chases scale by making models bigger, Liquid AI bets on scale by distribution — pushing intelligence outward, not upward.


Glossary

Term Definition
Agentic AI AI systems composed of agents (autonomous modules) that act, reason, or intervene toward goals, often via tools, memory, or decision logic.
Foundation model A large, general-purpose AI model (e.g. GPT, PaLM, Claude) typically pre-trained on large datasets and then adapted for specific tasks.
Parameter / parameter count The number of internal weights in a model; more parameters often (though not always) mean greater capacity.
Retrieval-Augmented Generation (RAG) A technique where a model augments its reasoning by fetching relevant external documents or data to ground responses.
Inference latency The time delay between input and output in a model performing its computation.
Edge / on-device inference Running model computations directly on the local device (phone, sensor, IoT) rather than in remote cloud servers.

Source: What if we’ve been doing agentic AI all wrong? MIT offshoot Liquid AI offers new small, task-specific Liquid Nano models — VentureBeat ([Venturebeat][1])

[1]: https://venturebeat.com/ai/what-if-weve-been-doing-agentic-ai-all-wrong-mit-offshoot-liquid-ai-offers “What if we’ve been doing agentic AI all wrong? MIT offshoot Liquid AI offers new small, task-specific Liquid Nano models VentureBeat”