Quick Facts
- Liquid AI released LFM2.5-230M on June 25, 2026, a 230-million-parameter model trained on 19 trillion tokens and built for on-device AI workflows.
- The model scored 22.51 on the CaseReportBench data extraction benchmark, beating Alibaba’s 800M-parameter Qwen3.5-0.8B (13.83) and Google’s Gemma 3 1B (2.28).
- Free for companies under $10 million in annual revenue; larger enterprises must sign a paid agreement.
Liquid AI, the MIT spinout valued at roughly $2 billion, released a new foundation model on June 25 that weighs in at just 230 million parameters but outperforms models with up to four times as many on structured data extraction tasks.
The model, called LFM2.5-230M, is designed for on-device and edge deployments where cloud access is limited or undesirable. It runs at 213 tokens per second on a Samsung Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5, both on CPU.
Benchmark Numbers
On the BFCLv3 tool-use benchmark, LFM2.5-230M scored 43.26. IBM’s Granite 4.0-350M scored 39.58. Google’s Gemma 3 1B IT scored 16.61.
The gap widens on data extraction. Using the CaseReportBench dataset, LFM2.5-230M posted 22.51. Alibaba’s Qwen3.5-0.8B, which has 800 million parameters, scored 13.83. Google’s Gemma 3 1B scored 2.28.
Architecture
The model runs on Liquid AI’s LFM2 architecture, a hybrid system that combines gated short-range convolutions with grouped-query attention. The design avoids the quadratic memory scaling that burdens standard transformer architectures, making it practical on constrained hardware.
LFM2.5-230M has 14 layers, a 32,768-token context window, and a vocabulary size of 65,536. It was pre-trained on 19 trillion tokens across 10 languages, including English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, and Spanish. The knowledge cutoff is mid-2024.
Post-training followed three stages: supervised fine-tuning with distillation from Liquid’s larger LFM2.5-350M model, direct preference optimization, and multi-domain reinforcement learning.
What It Cannot Do
Liquid AI is direct about the model’s limits. LFM2.5-230M is not built for advanced math, coding, or creative writing. The company describes it as a tool for lightweight data extraction pipelines and autonomous edge systems, not a general-purpose reasoning engine.
Deployment and Licensing
The base and post-trained models are available on Hugging Face with day-one support for llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX. Liquid AI also demonstrated the model running on a Unitree G1 humanoid robot, where it acted as a natural-language control layer on an onboard NVIDIA Jetson Orin chip, decomposing instructions into tool calls via NVIDIA’s SONIC framework.
The licensing structure is dual-use. Individuals and companies generating less than $10 million in annual revenue can use it for free. Companies above that threshold must sign a paid enterprise agreement.
Why It Matters for Builders
For software teams building data pipelines, document processing tools, or edge AI products, LFM2.5-230M offers a narrow but well-defined value proposition: fast, accurate structured data extraction that can run locally without API calls or cloud costs.
Liquid AI has raised $297 million in total funding, with AMD Ventures leading its $250 million Series A. The company’s architecture research traces back to MIT’s Computer Science and Artificial Intelligence Laboratory, where co-founder Ramin Hasani studied continuous-time dynamical systems as alternatives to fixed-weight neural networks.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.