Liquid AI releases LFM2.5-350M small model
Liquid AI has released LFM2.5-350M, a 350M-parameter model aimed at tool use and data extraction on edge devices, and says it outscores models more than twice its size on several benchmarks.
Liquid AI announced LFM2.5-350M, an updated version of its 350M-parameter model. The company extended pre-training from 10T to 28T tokens and added large-scale reinforcement learning. The base model (LFM2.5-350M-Base) and the post-trained model are available on Hugging Face, Liquid's LEAP platform and its Playground.
Liquid's own evaluations show the largest gains over LFM2-350M in three areas: instruction following (IFBench, 18.20 → 40.69), data extraction (CaseReportBench, 11.67 → 32.45) and tool use (BFCLv3, 22.95 → 44.11). The company says the model beats Qwen3.5-0.8B, Gemma 3 1B IT and IBM's Granite 4.0 350M variants on most of the ten benchmarks it ran. There are exceptions. Qwen3.5-0.8B Instruct scores much higher on MMLU-Pro (37.42 vs. 20.01), and its Thinking variant leads on BFCLv4. Liquid says the model is not recommended for math, code or creative writing.
On throughput, Liquid reports a peak of 40.4K output tokens per second on a single H100 SXM5 using SGLang 0.5.9. The test used 1,024 input tokens and up to 256 output tokens in BF16. Liquid says that works out to more than 3.5B tokens per day. It also says CPU inference is significantly faster than Granite-4.0-H-1B and Qwen3.5-0.8B but gave no figures in the material reviewed.
The model has day-one support for llama.cpp, MLX, vLLM, SGLang, ONNX and OpenVINO. Liquid named Zetic, RunAnywhere and Mirai as new ecosystem partners. It also said Distil Labs fine-tuned the model for multi-turn tool calling and reached over 95% accuracy, though no details of that test were given.