Liquid AI on Oct. 27 released LFM3, a family of six open-weight models in the 2-billion-parameter class that handle text, vision, audio, combined "omni" inputs and decision tasks, all built on a shared LFM3-1.7B-Base backbone.
The release is aimed at teams that want AI running on their own hardware. Liquid said in its launch blog that every model can be fine-tuned on a single GPU. The benchmarks behind the launch are company-reported, and VentureBeat found no independent evaluations as of Oct. 29.
What's in the family
Liquid's model catalog lists LFM3-1.7B for text, LFM3-VL-2B for vision and video, LFM3-Audio-2B for speech in and out, and LFM3-Omni-2B, which takes in text, images, video and audio and responds in text and speech. Two decision models, d1-2B and d1-Omni-2B, round out the set.
According to its Hugging Face model card, the VL model is the first in Liquid's LFM line to accept video. The Audio model handles 18 languages and relies on a tokenizer the company calls LiquidCodec. The Omni model produces 24 kHz speech and has a 32K context window, meaning it can consider roughly 32,000 tokens of input at a time.
The decision models sit on top of the others: d1-2B is built on LFM3-VL-2B, and d1-Omni-2B on LFM3-Omni-2B. Liquid said a single API call returns calibrated probabilities across a fixed set of outcomes, and that the models can "answer many questions in a single forward pass."
Size, speed and benchmarks
The text model has 1.73 billion parameters and a 256K context window. "The text model fits in about 1 GB at 4-bit precision," the Liquid AI team wrote in its Hugging Face launch blog. That is small enough for a phone.
On Liquid's own 29-benchmark average, LFM3-1.7B scored 49.37. Qwen3.5-9B came in at 47.24, Qwen3.5-4B at 39.27, MiniCPM5-2B at 39.19, Gemma4 E4B at 31.74 and Liquid's earlier LFM2.5-2.6B at 31.55. The company said LFM3-1.7B beats that predecessor with 35% fewer parameters.
Agentic tasks, where a model carries out multi-step work such as navigating websites, drove much of the gap. VentureBeat reported that on WebArena-Verified, LFM3-1.7B scored 38.42 against 28.94 for Qwen3.5-9B and 3.02 for LFM2.5-2.6B. On AutomationBench the figures were 41.29, 15.10 and 0.94.
Bigger models still hold the edge in other areas. On GPQA Diamond, a graduate-level science test, LFM3-1.7B scored 63.37 to Qwen3.5-9B's 80.40. On RULER 128K, which probes long-context recall, the split was 65.56 to 87.11, and on SWE-bench Verified, a coding benchmark, it was 43.40 to 51.20. On The Wire reported that Qwen3.5-9B leads on 15 of the 29 individual tests.
Liquid did not dispute the point. "Larger models still win on agentic coding and knowledge, where size matters most," the company said in a written response to VentureBeat. The model card goes further, saying the text model "is not recommended for agentic coding or complex multi-step reasoning due to its small size."
For vision, Liquid said LFM3-VL-2B averaged 71.4 across 62 tests, roughly level with Gemma 4 12B at 71.7.
Edge performance claims
Liquid said its audio stack reaches a time to first token as low as 50 milliseconds. It also said LiquidCodec generates speech about 87 times faster than real time on a single CPU thread of an Apple M4 Max.
In llama.cpp tests at 32K context, the company said, LFM3-1.7B was the fastest of the models compared on an AMD Ryzen AI Max+ 395, an Apple M5 Max and an iPhone 17 Pro. On a single NVIDIA B300 handling 1,024 concurrent requests (BF16, SGLang), it cited capacity of 3.6 billion tokens per day. All of these figures are company-reported.
Availability, pricing and the license
Weights are on Hugging Face in BF16, FP8, NVFP4, GGUF, MLX and ONNX formats. Liquid's models page lists support for Transformers, vLLM, SGLang, llama.cpp, MLX, Ollama and LM Studio.
Hosted API pricing for LFM3-1.7B, as listed Oct. 29, is $0.08 per million input tokens, $0.02 per million cached input tokens and $0.20 per million output tokens. The vision, audio and omni models cost more. VentureBeat noted that the hosted API lists a 128K context limit, against 256K for the downloadable text model.
The catch is the license. The LFM Open License v1.0 is Apache-based and royalty-free, but Liquid's license page limits free commercial use to companies with less than $10 million in annual revenue. Larger enterprises need a paid license. On The Wire contrasted that with the fully permissive Apache-style licenses some competitors use.
Partners and the company
Liquid named Ambarella, AMD, Arm, Intel, NVIDIA, OpenRouter, Qualcomm and Vercel as launch partners or platforms. Mercedes-Benz and Shopify were cited as pre-release testing and evaluation partners.
Liquid AI is an MIT CSAIL spin-out led by CEO Ramin Hasani, CTO Mathias Lechner and CSO Alexander Amini, with Daniela Rus as a co-founder. It raised a $37.5 million seed round in December 2023, followed by a $250 million Series A led by AMD Ventures on Dec. 13, 2024. Reporting placed the valuation from that round above $2 billion.
Hardware enablement is the next step. Liquid said Ambarella support for LFM3 is in progress, building on LFM2.5-VL, which already runs on the chipmaker's silicon.
