LiquidAI's LFM2.5-2.6B runs agentic AI on edge, beating 4x larger rivals

LiquidAI's LFM2.5-2.6B is a compact, edge-first agentic model that punches far above its weight. Pre-trained on roughly 34T tokens with a mid-training context extension to 128K, the model's post-training pipeline is notably sophisticated: it runs two rounds of supervised fine-tuning weighted toward agentic data, then trains domain-specific teachers (math, code, tool use) and distills them back into a single student using multi-domain on-policy distillation. The final stage—Agentic Reinforcement Learning—executes multi-turn RL inside real agent harnesses like OpenClaw and Hermes Agent, with a harness proxy that treats these environments as black boxes while capturing token-level trajectories for training. This focus on real-world compatibility is what sets LFM2.5-2.6B apart from models that only perform well in synthetic benchmarks.

On evaluation, LFM2.5-2.6B delivers standout results against models nearly 4x its size. It tops every instruction-following benchmark, including IFBench (59.17), Multi-IF (80.07), and IFStruct (85.49), and leads on tool use with ToolSandbox at 77.83, beating Qwen3.5-9B (76.44) and far surpassing both Gemma variants. It even edges out larger models on agentic tasks like Claw-Eval and BrowseComp+, and on knowledge benchmarks, though coding remains a weak spot—LiquidAI suggests reaching for a larger model for code-heavy tasks. Inference is equally impressive: with efficient LFM2 architecture, the model decodes at 220 tokens/s on an M5 Max, 113 tokens/s on a Ryzen AI Max+ 395, and even 30 tok/s on a phone, all while staying under 2.5GB memory. On a single H100, it hits nearly 15K output tokens per second, roughly 1.3B tokens per day. Day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX means developers can deploy it immediately for high-volume, privacy-sensitive agent workloads—no cloud bill required.