Liquid AI’s tiny on-device model is rewriting agentic AI economics
Liquid AI just shipped LFM2.5-2.6B, a 2.7-billion-parameter model that plans, calls tools and completes multi-step tasks entirely on a phone, laptop, PC or robot—no cloud required. The model sports a 131 072-token context window, a 128 000-token vocabulary and open weights under the lfm1.0 license, all available today on Hugging Face in GGUF, MLX and ONNX formats.
From pocket to fleet in one download
Deployment spans every tier of the stack. Solo developers can pilot on existing laptops—the model delivers about 220 tokens per second on an M5 Max with under 2.5 GB RAM. Mid-market teams can self-host on a single Nvidia H100 SXM5, handling roughly 1.3 billion tokens per day. Enterprises and OEMs can push the same weights to fleets via GGUF and ONNX, keeping prompts air-gapped. Liquid AI explicitly excludes agentic coding and heavy knowledge tasks, focusing instead on agentic workflows, tool use, data extraction, RAG and long-context parsing.
Training recipe turned into a Swiss-army knife
The architecture mixes 22 double-gated convolution blocks with 8 grouped-query attention layers across 30 total layers. Pre-training used roughly 34 trillion tokens; the team doubled the vocabulary to 128K by extending the existing tokenizer rather than retraining from scratch, then added a dedicated mid-training phase to push context to 128K. Post-training unfolds in four stages—two supervised fine-tuning rounds, domain expert reinforcement learning, multi-domain distillation and agentic reinforcement learning inside real harnesses such as Hermes Agent and OpenClaw.
Benchmarking against gemma-4-E2B-it (5.1B), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B) and Qwen3.5-9B (9.7B) shows LFM2.5-2.6B leading most instruction-following and tool-use tests, trailing only Qwen3.5-9B on BFCLv4.
LFM2.5-2.6B model page • Hugging Face base weights • GGUF weights
Why it matters
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.