Liquid AI, a company founded by former MIT computer scientists, has released its smallest language foundation model to date, LFM2.5‑230M. The model contains 230 million parameters and is explicitly designed for on‑device agentic workflows that perform data extraction, tool calling and autonomous edge system tasks.
Purpose and significance
According to Liquid AI, the model's small parameter footprint makes it feasible to deploy on smartphones, laptops, robots and other resource‑constrained devices. The company claims LFM2.5‑230M outperforms much larger models on selected benchmarks for data extraction — citing advantages over the 800‑million‑parameter Alibaba Qwen3.5‑0.8B (Instruct) and the 1‑billion‑parameter Google Gemma 3 1B in specific tasks.
The intended audience is developers and engineers building lightweight extraction pipelines and autonomous edge systems.
Architecture and performance
LFM2.5‑230M is built on the LFM2 framework rather than a standard transformer: it uses a hybrid approach that interleaves gated short‑range convolutions with grouped‑query attention to process information efficiently. This design aims to handle long contexts and sequential data on edge hardware without the quadratic memory costs typical of pure attention mechanisms.
The model supports a 32K context window for ingesting large documents or continuous telemetry streams. Liquid AI's published charts show a memory footprint under 400 MB while delivering prefill and decode speeds that compare favorably to models like Gemma 3 1B IT and Granite 4.0‑H‑350M.
Concrete device measurements reported by the company include a decode speed of 213 tokens per second on a Samsung Galaxy S25 Ultra with a Qualcomm Snapdragon Gen4 CPU, and 42 tokens per second on a Raspberry Pi 5. Internal benchmarking also indicates the GPU inference stack yields lower end‑to‑end latency than competing small models at all concurrency levels.
Why enterprises should care
The release frames the model as a cost‑effective alternative for routine enterprise data processing. Traditional rule‑based ETL scripts are brittle: small changes in document layout or a schema can break pipelines. The industry is shifting toward “AI ETL,” where machine learning infers mappings, detects schema drift and adapts automatically.
Using large flagship models for routine parsing tasks can be expensive — the article cites examples of large models charging $5.00 per million input tokens — so a lightweight on‑device extractor can reduce compute costs and latency by avoiding persistent cloud API calls.
Liquid AI positions LFM2.5‑230M as a specialized, efficient engine for automating repetitive formatting, data parsing and structured tool calls directly on local hardware.
Small‑model landscape and benchmarks
The AI industry in mid‑2026 has seen renewed focus on smaller models, though definitions of “small” vary. The company references larger 3‑billion‑parameter efforts like Weibo's VibeThinker‑3B and Google's Gemma 4 family, which aim to bring frontier capabilities to mobile and IoT devices.
By contrast, LFM2.5‑230M operates in a much smaller weight class: at 230 million parameters it is about one‑tenth the size of Google's smallest Gemma 4 model and Weibo's VibeThinker‑3B. Liquid AI acknowledges the model is not intended for heavy reasoning workloads such as advanced math, coding or creative writing.
In its targeted domains, however, the model reportedly outperforms larger alternatives. Liquid AI's benchmarks show an LFM2.5‑230M BFCLv3 (tool‑use) score of 43.26 versus 39.58 for IBM Granite 4.0‑350M and 16.61 for Google Gemma 3 1B IT. On CaseReportBench for data extraction, LFM2.5‑230M scores 22.51, outperforming Qwen3.5‑0.8B (Instruct).
The company argues these results demonstrate that while multi‑billion‑parameter models tackle advanced tasks, a well‑optimized 230‑million‑parameter model can be the superior choice for structured tool calls and efficient agentic pipelines on constrained hardware.
Advanced research applications
Because of its strength in tool calling, LFM2.5‑230M can act as a skill‑selection layer. Liquid AI demonstrated this by running the model on a Unitree G1 humanoid robot, executing entirely on the robot's onboard NVIDIA Jetson Orin compute module. The model translated free‑form instructions into structured multi‑step plans that invoked pre‑trained low‑level skills provided by NVIDIA's SONIC framework.
The base and fine‑tuned models are available immediately on Hugging Face, with day‑one native support across inference tooling including llama.cpp (GGUF), MLX, vLLM, SGLang and ONNX.
License and commercial terms
LFM2.5‑230M is released under the LFM Open License v1.0. Despite the word “Open,” the license is not OSI‑compliant and includes a dual‑use commercial restriction.
For independent developers, researchers and early‑stage startups, the license grants a perpetual, worldwide, royalty‑free right to reproduce, modify and distribute the model if original copyright notices are preserved and modifications are clearly disclosed. However, any legal entity generating $10 million or more in annual revenue is not permitted to use the model commercially under this license and must negotiate a paid enterprise agreement with Liquid AI to deploy it in production.
Liquid AI says the licensing approach aims to prevent large technology conglomerates from freely appropriating its intellectual property while still enabling broad grassroots adoption among developers.
Availability and closing notes
LFM2.5‑230M is available now on Hugging Face in the announced formats and with support across major inference ecosystems. Liquid AI frames the release as a pivot toward architectural efficiency rather than brute‑force parameter scaling, demonstrating that complex, multi‑step agentic workflows can run on constrained hardware without constant cloud connectivity.
