Liquid AI's 230M-Parameter LFM2.5 Beats Models Four Times Its Size at Data Extraction and Runs on a Raspberry Pi
The MIT spinout's smallest model yet pairs convolution and attention blocks to outscore billion-parameter rivals on data extraction while decoding at 42 tokens per second on a Raspberry Pi 5.
Editor's Note ·
- Correction:
- The article describes the model as using a 'hybrid 16-block layout.' The cited source (MarkTechPost) states the model 'has 14 layers total' — eight double-gated LIV convolution blocks plus six grouped-query attention blocks (8 + 6 = 14, not 16). The correct figure is 14 layers.
Overview
Liquid AI has released LFM2.5-230M, a 230-million-parameter foundation model that the company says outperforms substantially larger models on data extraction tasks while running on hardware as constrained as a Raspberry Pi. According to Liquid AI, the model is built to “run anywhere,” from cloud GPUs to low-cost CPUs, and ships as open weights that developers can “Download, fine-tune, and deploy without restrictions.”
The model is the smallest entry in the company’s LFM2 family and arrives with same-day support across five inference runtimes, according to MarkTechPost.
What We Know
Liquid AI is a foundation-model company that, per its own profile, “Spun out of MIT CSAIL” and is led by founders Ramin Hasani (CEO), Mathias Lechner (CTO), Alexander Amini (CSO), and Daniela Rus.
LFM2.5-230M departs from the attention-only transformer design that dominates large language models. According to MarkTechPost, the model uses a hybrid 16-block layout in which “Eight are double-gated LIV convolution blocks” and “The remaining six are grouped-query attention (GQA) blocks,” a structure the company chose to “target fast CPU inference.” The same source reports a 32,768-token context window, a 65,536-token vocabulary, and a memory footprint of “293–375 MB.”
The model was “pre-trained on 19 trillion tokens,” according to both Liquid AI and MarkTechPost, with the latter noting a knowledge cutoff of mid-2024.
On the data-extraction benchmark CaseReportBench, LFM2.5-230M scored 22.51, ahead of Alibaba’s 800-million-parameter Qwen3.5-0.8B at 13.83 and Google’s billion-parameter Gemma 3 1B at 2.28, according to Liquid AI. On instruction-following, MarkTechPost reports an IFEval score of 71.71 for LFM2.5-230M against 59.94 for Qwen3.5-0.8B and 63.49 for Gemma 3 1B. Liquid AI describes the model as “surprisingly capable at tool use and data extraction tasks” despite its size, according to Liquid AI.
The efficiency claims center on edge hardware. On a Galaxy S25 Ultra, the model reaches a decode speed of “213 tok/s,” and on a Raspberry Pi 5 it sustains “42 tok/s” decode, according to Liquid AI. Day-one deployment is supported across “llama.cpp, MLX, vLLM, SGLang, and ONNX,” per MarkTechPost.
What We Don’t Know
The benchmark figures are vendor-reported and have not been independently reproduced. CaseReportBench measures a narrow class of structured data-extraction tasks, so the model’s strong showing there does not necessarily transfer to general reasoning or open-ended generation. Liquid AI publishes broader scores, including a GPQA Diamond result of 25.41, that place the model well below frontier systems on harder reasoning, according to Liquid AI.
The license terms beyond the open-weight grant are not fully detailed in the primary announcement; MarkTechPost lists the license as lfm1.0 without enumerating its conditions.
Analysis
LFM2.5-230M is a bet that on-device workloads such as document parsing and tool-calling reward architectural efficiency over raw scale. By interleaving convolution and grouped-query attention rather than relying on attention alone, the model sidesteps the memory growth that makes large transformers impractical on phones and single-board computers. If the data-extraction results hold up under independent testing, the release strengthens the case that sub-billion-parameter models tuned for specific agentic tasks can displace larger general-purpose models at the edge.