Liquid AI's 2.6B LFM2.5 runs 30 tok/s on phones with 128K context

Liquid AI today unveiled LFM2.5-2.6B, a compact 2.69B-parameter model aimed squarely at local, device-side agentic workloads. The model supports 128K context and native tool calling, and was post-trained specifically for multi-step agent workflows. The official Q4_K_M GGUF weighs in at just 1.67GB and is already compatible with llama.cpp, with vendor-reported speeds of 30 tok/s on a phone, 113 tok/s on a Ryzen AI Max+ 395, and 220 tok/s on an Apple M5 Max while staying under 2.5GB memory. These are vendor benchmarks, so real-world numbers still need independent verification.

On benchmarks, LFM2.5-2.6B is surprisingly competitive for its size: it beats Qwen3.5-9B on ToolSandbox (77.83 vs 76.44) and IFBench (59.17 vs 56.47), but falls behind on BFCLv4 (56.88 vs 60.13) and more clearly on LiveCodeBench (59.41 vs 69.86). Liquid's own model card warns it is not recommended for agentic coding. The realistic use case is as a cheap, private worker agent handling extraction, searches, and repetitive tool calls locally, with a larger model stepping in only for planning. The 128K claim also needs practical testing—supporting long context and actually running it on a phone with KV cache overhead are different things. Community testing on Android, older laptops, and mini-PCs is needed.