Liquid AI

Yeah, so they want to stake in the software layer that's actually going to power all these efficient on-device models.

And when we talk about efficient, we have to talk about Liquid AI.

They released LFM 2.5 24BA2B.

Which is a terrible name, but a brilliant piece of engineering.

Yeah, the money is violently diverging right now into two absolute extremes.

On one side, you have Google's Gemini Argon, massive centralized frontier models built for high margin enterprise logic.

And on the exact opposite side, you have Liquid AI and Meta's Spark Open.

Hyper-efficient, decentralized, on-device models optimized for privacy and speed.

And here we are today with models that are...

Today our guest is Jeffrey Lee, COO at Liquid AI and founder of Reza Health.

Liquid AI is betting that AI's future is smaller, faster, and running directly on your devices, not just in massive data centers.

After scaling companies like Scale AI and GatherTown, Jeffrey is now guiding their elite team to deploy real-world AI for customers like Mercedes.

I met with the Liquid team when they were only a couple months old before they had any models.

We're the founders of Phelonic, and we build engineering teams that actually function for startups by combining agile training with psychology, neuroscience, and semantics.

This is our show about what really makes teams tick.

Our guest today is Jeffrey Lee, COO of Liquid AI, the foundation model company spun out of MIT that's betting the future of AI is smaller, more efficient, and running on the devices in your pocket, your car, and your earbuds, not just in massive data centers.

This is huge.

Liquid AI just launched a fascinating open model.

It runs complex autonomous agents locally on high-end laptops.

The privacy implications there are massive.

If you run agents locally, your corporate data never leaves your machine.

Exactly.

And Liquid AI's model actually beat Alibaba's much larger Quinn 3.6 model.

It won five out of eight agent capability tests, all while running locally.

The pricing there is set at $2 per million input tokens and $10 per million output tokens.

So the, the, the pricing wars are definitely heating up, but, you know, the architectural shifts are, are even more significant, and MIT spin-out Liquid AI just released an open model called LFM 2.5-24B-A2B.

Quite a mouthful.

Yeah, but it is specifically designed to run powerful agentic workflows entirely locally on high-end laptops.

They want a centralized system to log AI containment failures, and they're demanding that labs keep any hacking-capable models strictly offline by default.

This is the ultimate juxtaposition.

On one hand, Liquid AI is engineering models specifically so you can run powerful agents locally.

On the other hand, the godfathers of AI are begging for a global containment grid.

Should this request go to the big model? Is this message safe? A decision model takes the question and the state, runs a single forward pass and returns a structured answer, like a calibrated probability, with zero output tokens.

Liquid AI released open weights for D1-3B and an experimental D1-OMNI-600M.

The 3B answers a question in 8 milliseconds on an RTX 4090.

Hugging Face released JEV27B-VL, which takes text and images and outputs calibrated probabilities.