Liquid AI released LFM2.5-2.6B on August 4, 2026, a 2.6B-parameter model trained for agentic workloads that can run entirely on-device. It is not positioned as a replacement for every large cloud model. The goal is to let phones, CPUs, and other edge devices handle planning, tool calling, and multi-step tasks locally.
The practical difference is that an agent does not need to call a cloud API for every step. Liquid AI says local models can reduce latency and data exposure while avoiding per-token inference charges. If the model is small enough, teams can parallelize background work on local hardware. Production deployment still has to address battery use, device security, model updates, tool permissions, and human takeover when a task fails.
The LFM2.5-2.6B training pipeline combines supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning. Liquid AI says the model was pretrained on roughly 34 trillion tokens, expanded its tokenizer to 128K, and received a 128K context-extension phase for long inputs and agent traces. The final training runs directly inside real harnesses and sandboxes such as Hermes Agent and OpenClaw, exposing the model to actual tools, system prompts, and interaction patterns.
The company’s comparisons cover STEM, instruction following, tool use, and agentic workflows. Liquid AI reports that LFM2.5-2.6B leads on every instruction-following benchmark and nearly every tool-use benchmark, trailing Qwen3.5-9B only on BFCLv4; larger models still lead on coding tasks. These are publisher-reported comparisons under specified settings, so they are useful selection signals rather than independent guarantees across hardware, quantization, and real workflows.
The speed figures show the intended deployment profile. In its tests, Liquid AI reports about 220 tokens per second on an M5 Max and 113 tokens per second on a Ryzen AI Max+ 395 while staying below 2.5GB of memory, with roughly 30 tokens per second on a phone. Such results depend heavily on hardware, runtime, quantization, prompt length, and concurrency. Teams should rerun their own representative tasks before committing to a device target.
The integration path is straightforward. The base and post-trained weights are available on Hugging Face, and the company suggests serving the local model behind an OpenAI-compatible endpoint before connecting it to harnesses such as Hermes Agent, OpenClaw, or Pi. Existing agent systems can keep their tool interfaces while changing the model route, but tool schemas, long-running recovery, and security policy still need to be tested.
LFM2.5-2.6B shifts model selection away from asking for the strongest model and toward matching task, latency, privacy, cost, and device constraints. A small local model can be easier to run continuously for offline processing or repetitive high-volume work, while larger models may remain better for difficult coding and reasoning. A sensible design is to route different task classes to different models and preserve permissions and human review for each class.
