Liquid AI Releases LFM2.5-2.6B On-Device Agent Model; 220 tok/s on M5 Max, Matches 4-10B Models

Liquid AI released LFM2.5-2.6B on August 4, 2026, a 2.6-billion-parameter model designed for on-device agentic workloads with no cloud dependency. The model runs at 220 tokens/second on Apple M5 Max, 113 tok/s on AMD Ryzen CPU, and even 30 tok/s on phones, all within 2.5 GB of memory. It was pre-trained on ~34 trillion tokens with a 128K context window and post-trained in four stages: supervised fine-tuning, expert specialization, multi-domain on-policy distillation, and agentic reinforcement learning inside live harnesses (OpenClaw, Hermes Agent).

LFM2.5-2.6B is competitive with models 4-10x larger on tool use, instruction following, and multi-step agentic tasks. On tool-use benchmarks (BFCLv4, ToolSandbox, Claw-Eval), it matches or beats Gemma 5-8B and Qwen 4.7-9.7B models. The architecture uses mostly short convolutions with selective attention layers (LIV convolutions), which maintain constant-size state per token and eliminate KV cache overhead—a key efficiency advantage over dense transformers. GPU inference reaches 15K output tokens/second on a single H100 at high concurrency.

For architects shipping agents: on-device agentic models eliminate per-token cost and enable privacy-by-default inference. LFM2.5-2.6B's training inside real harnesses (not synthetic traces) should improve actual tool-use compatibility. The 30 tok/s on phones opens robotics and embedded-AI use cases. The cost model flip—from token-spend constraint to local-throughput constraint—enables continuous background agents. Liquid's emphasis on convolution-based architectures over pure attention may signal where efficient edge models are headed.