AIBRIEF
The 2.5 GB model beats some 9B class systems on tool use, loses a public code benchmark to a 9 billion parameter rival, and carries a 'not for coding' disclaimer from its maker.
Liquid AI's LFM2.5-2.6B, a 2.6-billion-parameter hybrid of 22 convolution blocks and 8 grouped-query attention layers, runs tool-using agents on a laptop, a phone, or a CPU-only box without sending data to a cloud API. Liquid reports 220 tokens per second on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, with the model under 2.5 GB of memory and ~30 tok/s on a phone.
The model edges out Gemma 4-E2B-it, Gemma 4-E4B-it, and Qwen3.5-4B on most agentic tests in Liquid's benchmark table, and AlphaSignal reports it beat Qwen3.5-9B on ToolSandbox (77.83 vs 76.44). It loses LiveCodeBenchv6 to that same Qwen3.5-9B. Its AA-Omniscience score is -29.50, negative for every model in the table: "smaller is less bad," not good.
The blog's τ³-Bench Banking 5.67 is absent from BenchLM and llm-stats, and the model is not on the Berkeley Function Calling Leaderboard. An independent review of the older LFM2-2.6B flagged weak reasoning depth, a family-level floor rather than a 2.5 result.
Liquid's own disclaimer is the cleanest caveat: the model is "not recommended for coding or knowledge-intensive tasks." For local agent stacks that need browsing, API calls, and multi-step tool use on commodity hardware, the throughput numbers ship today; the 9B-class claims need an independent rerun.