Liquid AI Releases LFM2.5-2.6B: A 2.6B-Parameter Agent Model Built for On-Device Deployment

On August 4, Liquid AI released LFM2.5-2.6B on the Hugging Face blog, a 2.6B-parameter model built to power capable AI agents entirely on-device. The company says it supports tool calling and multi-step workflows, letting developers deploy agents on everyday hardware ranging from laptops to phones.

The positioning is explicit: agents that run locally keep data on the device and avoid a growing cloud inference bill. Liquid AI claims the model is competitive with models about four times its size on tool use, instruction following, and multi-step agentic tasks.

Behind the release is a substantial training pipeline. LFM2.5-2.6B was pre-trained on roughly 34T tokens, with a mid-training phase that extends the context window to 128K. Post-training turns the base model into an agent in four stages: two rounds of supervised fine-tuning weighted toward agentic data, per-domain specialist teachers, multi-domain on-policy distillation (MOPD) that merges the teachers into one student, and agentic reinforcement learning run inside real agent harnesses.

Notably, the agentic RL phase trains the model inside popular harnesses such as OpenClaw and Hermes Agent, so it learns to work across different tools, system prompts, and multi-turn environments, improving real-world compatibility rather than just benchmark scores.

In benchmarks, Liquid AI pitted the model against options roughly four times larger, including gemma-4-E2B-it (5.1B), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B). The 2.6B model tops every instruction-following benchmark in the comparison and leads on nearly all tool-use benchmarks, with only the 9.7B Qwen edging ahead on BFCLv4; it stays even with the Qwen models on agentic tasks and beats both Gemma models, while coding remains the one area where larger models keep a clear lead.

Speed is the headline metric. Liquid AI reports decode rates of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, all in under 2.5GB of memory, with phones still able to run agents at around 30 tokens per second. On GPUs, the model reaches nearly 15K output tokens per second at high concurrency, roughly 1.3B tokens per day on a single H100.

Day-one ecosystem support covers llama.cpp, MLX, vLLM, SGLang, and ONNX, and the model loads with transformers>=5.0.0; a browser demo is live. For teams that want agents on laptops, phones, and edge devices, LFM2.5-2.6B offers a low-cost local path — the questions worth watching are how quickly it gets adopted in real on-device workflows and whether Liquid AI follows with larger variants.