Liquid AI squeezed six models into 2 billion parameters, but big companies may not deploy them commercially

Liquid AI released the LFM3 family on October 27: six models of around 2 billion parameters, all built on a single LFM3-1.7B-Base. At 4-bit precision the text model fits in about 1 GB of memory, and the company says it beats the previous LFM2.5-2.6B on agentic tasks with 35% fewer parameters.

One backbone, six specialists and a text model in 1 GB

LFM3-1.7B-Base spawns the text model LFM3-1.7B, the vision model LFM3-VL-2B, the audio model LFM3-Audio-2B, LFM3-Omni-2B, which handles vision and audio together, and two decision models, d1-2B and d1-omni-2B. The base was trained on 22 trillion tokens. The final 3.5 trillion carried a larger share of agentic traces and tool calls, which also stretched the context window to 256K tokens.

The company designed the architecture by measuring it directly on 14 device and backend combinations, from phones and a Raspberry Pi 5 to the Qualcomm Hexagon NPU. The result has 24 layers: 13 short convolutions, 5 sliding-window attention layers and 6 global attention layers. Per parameter, Liquid AI says it runs about 25% faster than LFM2-1.2B with 8K tokens in context.

The company also gives concrete numbers. LFM3-1.7B scores 38.42 on WebArena-Verified, against 3.02 for the previous LFM2.5-2.6B. LFM3-VL-2B, at 1.9 billion parameters, raises the average across 62 vision benchmarks from 57.6 to 71.4, which the company says puts it on par with Gemma 4 12B. The audio model speaks 18 languages and can return its first token in as little as 50 ms. The decision model d1-2B replaces d1-3B and scores 75.5 against 64.0 on the company's own ImajevBench.

Teams with their own hardware get more models per gigabyte of memory

Liquid AI is betting on the opposite of what the big labs do. Instead of one general model that does everything and runs in the cloud, it offers small specialists that a team picks by input type, fine-tunes on a single GPU and runs wherever the application lives. For device makers, robotics teams or companies that cannot send data out, that means one toolkit for voice control, reading receipts and triaging a support ticket.

The text model is also explicitly trained for agent harnesses such as Hermes Agent, Pi and OpenClaw. Liquid is aiming at developers who want a local agent on a laptop without an API provider account. It adds ten ready-made fine-tuning notebooks, from turning receipts into JSON to picking the right tool out of 108.

Compared with d1-3B, which we covered three weeks ago, Liquid now backs its decision models with numbers. They still come from four in-house test suites, not a public comparison.

The open-weights license stops at $10 million in revenue

On Hugging Face the models are labeled open-weight, but the LFM Open License v1.0 makes commercial use conditional on a company staying under $10 million in annual revenue. A larger company can download and test the weights, but this license does not cover its commercial deployment. The only exception is non-profits using them for non-commercial or research purposes.

The second limit is the numbers. Nearly all comparisons come from Liquid AI and were measured on its terms. The match with Gemma 4 12B is an average across 62 benchmarks, not a guarantee that the small model handles the same tasks in production. The WebArena-Verified jump from 3.02 to 38.42 mostly shows how weak the predecessor was at agentic work.

Independent tests and the first fine-tuned deployments will show how much holds

The first signal will come from independent benchmarks and the local-agent community. If LFM3-1.7B holds up in other people's harnesses, not just in the vendor's notebook, that will be a stronger argument than any blog table.

The second signal is robotics, which Liquid presents as its flagship example. The company reports 96.2% success on the LIBERO suites and action generation under 50 ms on a Jetson AGX Orin. Only deployments outside the lab will show whether a small specialist can carry work that a large general model does in the cloud today.

Lilith's verdict

Liquid AI is handing out six small specialists that fit in a pocket, the text one in a single gigabyte. Anyone who grows past $10 million in revenue will discover it was a free sample at the pharmacy, and the full pack is waiting at the checkout.