Liquid AI's LFM2.5-2.6B Matches Models Three Times Its Size on Agentic Tasks
Liquid AI released LFM2.5-2.6B on August 4, 2026 — a 2.69B-parameter model built from the ground up for on-device agentic workloads. The model ships as a standalone binary designed to run on consumer hardware without cloud dependencies, and it uses a hybrid architecture that Liquid AI says is fundamentally different from the transformer stack almost every other frontier model is built on.
The architecture mixes 22 double-gated short convolution blocks with 8 grouped query attention (GQA) layers across 30 layers total. The convolution layers handle sequence processing without the quadratic memory overhead that makes standard attention expensive at longer contexts, while the GQA layers handle the parts of the workload that benefit from explicit attention. The result is a model that runs under 2.5 GB of memory, supports a 128K-token context window, and reaches approximately 220 tokens/s on Apple M5 Max, 113 tokens/s on AMD Ryzen AI Max+ 395, and around 30 tokens/s on mobile hardware.
On the benchmarks Liquid AI published, the model is competitive with significantly larger models in its specific target domain. On ToolSandbox — which tests realistic tool-use workflows — LFM2.5-2.6B scored 77.83, edging out Qwen3.5-9B's 76.44 despite being roughly a third of the parameter count. On IFBench (instruction following), it scored 59.17 versus Qwen3.5-9B's 56.47. Performance diverges on tasks where raw scale matters more: on LiveCodeBench, Qwen3.5-9B (69.86) substantially leads LFM2.5-2.6B (59.41), which is broadly consistent with Liquid AI's framing — this model is designed to win at agentic orchestration and tool calling, not at general code generation or knowledge-intensive reasoning.
Both the instruction-tuned LFM2.5-2.6B and the LFM2.5-2.6B-Base checkpoint for fine-tuning are available on Hugging Face under the LFM Open License v1.0. The release fits into a broader market shift toward models that run locally and handle multi-step agent loops without requiring an API call per step — a segment that has accelerated alongside the agentic AI push from companies like Gemini Managed Agents and KAT-Coder-V2.5, but from a different angle: Liquid AI is betting on on-device execution and privacy-preserving inference rather than cloud-scale orchestration.
