Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, and PCs, with the open weights posted to Hugging Face on August 4, 2026. The pitch is simple: full AI agents, not just a chatbot, running locally where the marginal cost of each run is essentially zero.
Try It: Run a Local Agent This Afternoon
Both checkpoints are open on Hugging Face, the instruction-tuned LiquidAI/LFM2.5-2.6B and the base model. It runs through transformers, llama.cpp, MLX, vLLM, SGLang, and ONNX, so you can pull it into an existing local setup without new infrastructure. On an Apple M5 Max it hits 220 tokens per second, on a smartphone around 30, and it fits in under 2.5 GB of memory. Point it at a tool-calling harness and you have a background agent that can run web-search and multi-step automations locally, no API bill.
Why It Matters for Creators
On-device inference changes the economics of building with agents. Because there is no per-token cloud cost, you can parallelize workloads on your own hardware and run background tasks that consume millions of tokens for free, with lower latency and data that never leaves the machine. For creators wiring AI into local pipelines, batch captioning, asset tagging, file organization, or a private research assistant, that removes both the recurring bill and the privacy exposure of shipping everything to a cloud API. It sits alongside the broader push to run capable models locally on Mac hardware.
Key Details
Model: LFM2.5-2.6B (instruction-tuned) and LFM2.5-2.6B-Base, from Liquid AI
Parameters: 2.6 billion, 128K token context, LFM2 edge-optimized architecture
On-device speed: 220 tok/s on M5 Max, 113 tok/s on AMD Ryzen AI Max+ 395, ~30 tok/s on phone; under 2.5 GB memory
Capabilities: Tool calling, multi-step planning, multi-turn reasoning; trained inside real agent harnesses
Benchmarks: Competitive with models up to 4x larger on instruction following and tool use (IFBench 59.17, Multi-IF 80.07, BrowseComp+ 26.89)
What to Do Next
Download the instruction-tuned checkpoint, load it into your existing local runtime, and give it a small tool-calling task you would normally route to a cloud model, such as searching and summarizing, or renaming a folder of files by content. Compare speed and quality against whatever hosted model you use now. If the on-device run holds up, move your high-volume, privacy-sensitive automations local and keep the cloud for the heavy reasoning jobs.
