Liquid AI released LFM3 on 27 October, a family of six open-weight models that each weigh in around 2 billion parameters. Four handle text, vision, audio, and a combined “omni” mode; the other two are decision models that return judgments instead of prose. All six grow from one shared backbone, LFM3-1.7B-Base, which Liquid says was trained on 22 trillion tokens, according to the company’s launch post on the Liquid AI blog.

The design target is a model you can fine-tune yourself and run without a data center. Liquid searched for its architecture with real hardware in the loop, and the test bench grew from a lone phone and a lone laptop in LFM2 to 14 combinations of device and software backend for LFM3. Among them: a Samsung Galaxy S26 Ultra, a OnePlus 15 and a Raspberry Pi 5, plus an Apple M4 Max, AMD Ryzen AI laptops and the Qualcomm Hexagon NPU. For a team that ships software to phones or laptops, that coverage is the real selling point.

The text model, LFM3-1.7B, is tuned to work inside agent frameworks including OpenClaw, Pi and Hermes Agent, and it reads up to 256K tokens of context. Liquid’s table puts its average across the agent, coding, knowledge and long-context tests at 49.37, ahead of Alibaba’s Qwen3.5-9B at 47.24 and far ahead of its own LFM2.5-2.6B at 31.55. The bigger models keep their edge where raw size counts. On SWE-bench Verified, a coding test, Qwen3.5-9B scores 51.20 to LFM3’s 43.40, and on the GPQA Diamond science quiz it leads 80.40 to 63.37. Every figure comes from Liquid’s own comparison, with no outside lab involved.

Speed and memory are where the small size pays off. On llama.cpp at Q4_0 with a 32K-token context, Liquid reports the text model generating text between 1.8 and 2.8 times as fast as the runner-up on all three test devices. Holding a full 128K-token context costs 1.6 GB of cache, where Qwen3.5-4B needs 4.3 GB.

The most concrete outside evidence comes from MacPaw, the Mac software maker. Its engineers pitted LFM3-1.7B versus Qwen3.5-2B across 65 routine agent jobs inside actual macOS setups. The Liquid model finished 64 percent of them to Qwen’s 48, used 58 percent fewer tokens per task, and got stuck in under 1 percent of attempts against 15. “LFM3 is the first sub-2B model we’ve tested that is ready for real agent work on a Mac,” said Ivan Petrukha, MacPaw’s LLM efficiency lead.

The vision model, LFM3-VL-2B, is the first in the line to accept video. Across 62 benchmarks Liquid scores it at an average of 71.4, level with Gemma 4 12B (71.7), a model more than six times larger, and just under GPT-5.6 Luna at 73.4. The weak spot is agentic work on screens and tools, where it scores 42.5 while Gemini 3.5 Flash Lite reaches 62.4.

The audio model supports 18 languages and can start answering as quickly as 50 milliseconds after you stop speaking. The two decision models, d1-2B and d1-Omni-2B, take a ticket, an invoice, a photo or a voice note plus a list of questions and answer every question in one pass, each with a probability attached. d1-2B uses a third fewer parameters than the d1-3B it replaces and scores 85.2 on Liquid’s Decision Bench against 83.6.

Liquid also shows what a short fine-tune can do. With about 3,000 public photos of wind turbine blades and a few GPU hours on a single A100, the vision model went from identifying defect types in under 10 percent of images to 67 percent accuracy. Its detection score of 0.11 is still low in absolute terms, though far above the 0.01 Liquid measured for GPT-6 Luna.

The weights are on Hugging Face, with hosted access through the Liquid AI API and a Playground. Anyone paying cloud rates for routine agent loops can now repeat MacPaw’s comparison on their own tasks with free weights, and that result will say more than any vendor table.

Reported by the Liquid AI blog on 27 October 2026.