Liquid AI Releases 2.6B On-Device Agent Model

Liquid AI put out LFM2.5-2.6B this week, a model built to run agentic workloads entirely on local hardware, no server round-trip required. At 2.69 billion parameters, it fits on phones, laptops, PCs, and robots, with a 131,072-token context window and a 128,000-token vocabulary behind it. Pre-training drew on roughly 34 trillion tokens, and the company shipped two versions: LFM2.5-2.6B-Base for developers who want to fine-tune, and a post-trained LFM2.5-2.6B tuned specifically for planning, tool calls, and multi-step tasks.

The pitch is straightforward: since inference happens on the device itself, prompts never touch a third-party API, and each run costs close to nothing to execute. Liquid AI says the model holds its own against systems roughly four times its size on tool-use and instruction-following benchmarks, a claim that matters most for anyone trying to cut inference costs or keep data off outside servers entirely. Both checkpoints are live on Hugging Face under the lfm1.0 license, with weights available in native, GGUF, MLX, and ONNX formats. That range of formats is the detail worth sitting with before getting into how it actually performs.