Liquid AI’s new model LFM2.5-2.6B is positioned as a lightweight alternative for edge AI, promising to run on devices ranging from smartphones to a Raspberry Pi without cloud or GPU assistance.
Technical specs and performance claims
The model contains 2.6 billion parameters and supports a 128,000‑token context window. It ships with native tool‑calling capabilities and is available both as a post‑trained model and a base checkpoint for fine‑tuning on Hugging Face. Day‑one support includes inference stacks such as llama.cpp, MLX, vLLM, SGLang, and ONNX, which should make deployment across consumer hardware, enterprise servers, and embedded systems feasible.
Liquid AI‑reported benchmarks show decoding speeds of roughly 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, each using under 2.5 GB of memory. A smartphone can manage about 30 tokens per second through the Apollo mobile app, while a single Nvidia H100 GPU reaches close to 15,000 output tokens per second under sustained load. These figures have not been independently verified.
Focus on agentic workloads
Liquid AI states the model excels at high‑volume, well‑defined tasks such as document management, calendar automation, and continuous background routines. It is less suited for heavy coding workloads, which still benefit from larger frontier models.
In practice, the model’s small footprint means it can be embedded in environments where latency, privacy, or connectivity are critical—think vehicles, robotics, or regulated industries that cannot send sensitive data to the cloud.
Related: Tactile Technologies honors Womandla Women in STEAM
From a user standpoint, this shift could make AI assistants feel less like remote services and more like local utilities, reducing reliance on expensive token‑based inference.
Licensing and enterprise considerations
LFM2.5-2.6B is released under a custom open‑weights license, similar to other recent models, and may require careful review by enterprise legal teams.
Legal teams will likely scrutinize the license terms, especially for companies that handle regulated data and need to ensure compliance when deploying locally.
One practical implication is that enterprises can now consider deploying a single model across multiple agentic functions, swapping out the surrounding harness or tools rather than retraining new models for each use case. This could lower the total cost of ownership and simplify maintenance.
Edge deployments become more viable.
