Liquid AI LFM3: Six Open 2B Models Built for Devices

Liquid AI LFM3 is a bet that small, specialized models beat one big generalist for a lot of real work. Released on October 27, LFM3 is a family of six open-weight models around 2 billion parameters, covering text, vision, audio, omni-modal input and two "decision" models. All of them branch from one shared LFM3-1.7B-Base, and Liquid says each can be fine-tuned on a single GPU and run on phones, laptops or private servers.

What actually happened

Liquid published the release on its own blog and on Hugging Face, with weights available on Hugging Face, its Playground and its API. The base model was pre-trained on 22 trillion tokens. The last 3.5 trillion were a mid-training stage on curated data with more agent traces and tool calls, which is also where the context window was extended to 256K tokens.

The six models are LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B for speech, LFM3-Omni-2B for mixed audio and vision, and two decision models, d1-2B and d1-omni-2B, which return structured, calibrated answers instead of chat. The decision line follows the earlier d1 models we covered in Liquid AI d1.

The architecture is what Liquid calls a tri-hybrid. Of 24 layers, 13 are gated short convolutions, 5 use sliding-window attention over 256 tokens, and 6 use global attention. Liquid tuned the design with an automated search that tested 14 device and backend combinations, from a Samsung Galaxy S26 Ultra and a Raspberry Pi 5 to Apple and AMD laptops and a Qualcomm NPU.

How Liquid AI LFM3 stacks up

Here the details matter, and all scores are Liquid's own. VentureBeat's table shows LFM3-1.7B scoring 38.42 on WebArena-Verified against 28.94 for Qwen3.5-9B, and 41.29 on public AutomationBench against 15.10. On BFCL V4 function calling it scored 71.39 versus 62.83.

It loses elsewhere. Qwen3.5-9B leads on MCP-Atlas, SWE-bench Verified, GPQA Diamond and the 128K RULER long-context test. Liquid told VentureBeat that "larger models still win on agentic coding and knowledge", and its documentation says the text model is "not recommended for agentic coding or complex multi-step reasoning due to its small size".

The vision model posts the boldest claim. Liquid says LFM3-VL-2B, at 1.9 billion parameters, raised its average across 62 vision benchmarks to 71.4 from 57.6 for the older LFM2.5-VL-3B, which it puts on par with a model more than six times larger. Again, that is company-reported. Liquid names Mercedes-Benz and Shopify as pre-release testing partners, but VentureBeat notes there is no confirmed production rollout at either company yet.

A 2B model that clicks through a browser well is useful. A 2B model that claims to do everything is marketing.

The fine print on licensing and memory

VentureBeat flags two catches. First, the downloadable weights use the LFM Open License v1.0, and businesses above $10 million in annual revenue need a separate commercial agreement. Second, the "1 GB" figure is model size only. Liquid's own speed tests at a 32K input reported 1.9 GB to 2.75 GB of memory in use on an iPhone 17 Pro, an Apple M5 Max and an AMD system. The hosted API also lists a 128,000-token context limit, half the downloadable model's window.

This is a very different pitch from bigger open releases like the Nanbeige4.3-33B model. Liquid is not chasing the top of leaderboards. It wants to own the narrow, repeatable jobs that run on devices companies control.

What to do as a reader

If you build agents, test Liquid AI LFM3 on one bounded task, such as extracting a field from documents or routing tickets, and measure failure rates, not just averages. Small models break in different ways than big ones.

Check the license before shipping anything commercial, especially if your company is above the $10 million revenue line. Plan memory around real context lengths, not the headline 1 GB.

Ignore any crypto token claiming to be tied to Liquid AI or LFM3. Liquid has announced nothing of the kind. Nothing here is investment advice.

Sources: Liquid AI blog, Hugging Face blog and VentureBeat.