Liquid AI has released LFM2.5-2.6B, a small model aimed at local and on-device agent workloads.
The Hugging Face article published August 4 describes the model as a 2.6B-parameter release with tool calling, multi-step workflows, a 128K-token context window, and day-one support across llama.cpp, MLX, vLLM, SGLang, ONNX, Apple MLX-LM, and NVIDIA NIM.
Liquid AI also claims the model reaches 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395. Those are company-published speed claims and should be treated as a starting point for local testing, not as general proof that every device workflow will feel fast.
Small models are moving toward agent loops
The local-model story used to be mostly about private chat, offline summarization, and cheap batch work. LFM2.5-2.6B is framed around a wider target: agents that call tools, follow multi-step instructions, and work with long context without requiring a cloud endpoint.
That matters because many useful agent tasks are small but sensitive. A local assistant may inspect a personal file, route a local script, summarize a private document, or automate a device workflow. In those cases, a small model does not need to beat a frontier model. It needs to be reliable enough for the task and cheap enough to run close to the data.
The 128K context claim is especially relevant for local use. Long context can make a small model more useful for repository scans, support logs, local notes, and document batches. It also increases the need to test retrieval quality, instruction following, and failure modes under realistic context pressure.
Compatibility is part of the launch
Liquid AI’s support list is the other important piece. Models meant for local deployment rise or fall on the surrounding runtime ecosystem. If a model works cleanly across common inference stacks, developers can compare speed, quantization, memory use, and quality without waiting for one vendor endpoint.
That makes LFM2.5-2.6B a practical benchmark for edge-agent work. It is small enough to test quickly, but the claimed feature set points at real workflows rather than single-turn demos.
