Liquid AI unveils LFM3: six compact models from a single base, with in-house benchmarks and a restrictive license

On 27 October 2026, Liquid AI announced on its official blog the LFM3 family, a new series of models of roughly two billion parameters built to run locally on PCs and mobile devices. The whole lineup grows out of a shared foundation, LFM3-1.7B-Base, from which six variants were derived for text, vision, audio, omnimodal processing and specialized decision-making tasks. The text model, with 1.73 billion parameters and a 256,000-token context window, uses a 24-layer hybrid architecture that alternates gated short convolutions, sliding-window attention and global attention layers. Training started with 22 trillion tokens, followed by fine-tuning stages for context extension and instruction following.

In figures released by the company, the text model takes up about one gigabyte of memory with 4-bit quantization and, by the same measurements, needs a 1.6 GB KV cache at 128,000 tokens, versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. Company documentation reports an average of 49.37 points across nineteen benchmarks selected by Liquid AI, against 47.24 attributed to Qwen3.5-9B. The same table, however, shows the larger model staying ahead on scientific reasoning in GPQA Diamond (80.40 vs 63.37), long-context handling in RULER 128K (87.11 vs 65.56) and code generation in SWE-bench Verified (51.20 vs 43.40). According to Tech Times, at the time of publication these company-reported figures had not yet been reproduced by third parties.

The weights are published on Hugging Face and available through an API and dedicated playgrounds; according to Liquid AI, every model in the family can be fine-tuned on a single GPU. Under the LFM Open License v1.0, commercial use is free only up to US$10 million in annual revenue; above that threshold a custom commercial license is required, while research and non-profit use carry no financial conditions. Because of that threshold, the license is not open source in the OSI sense. On the growth of edge computing, Vinesh Sukumar, vice president at Qualcomm Technologies, said that "The next generation of AI will be more multimodal, personalized, and private", while Josh Hort, Senior Director at AMD, noted that "developers need models that can be deployed efficiently across the broad spectrum of AI PCs".

Fitting agentic tasks into a text model that, according to the company, takes about one gigabyte at 4 bits is careful engineering, shifting the emphasis from sheer parameter count to strict management of hardware resources. What remains to be measured is the gap between a vendor-picked set of benchmarks and real performance in everyday use, where the architecture's robustness will be judged on its ability to handle complex tasks across varied local setups. — Olya