On-device agents face an awkward trade-off: a model must be smart enough to plan and call tools, yet small and fast enough to run continuously on a laptop. Once privacy or offline use enters the picture, "data never leaves the device" becomes a hard requirement. Liquid AI's latest answer is LFM2.5-24B-A2B — the largest and, per the company's own wording, the final model in the LFM2.5 family, released October 13 with weights on Hugging Face the same day.

Architecture: 24B Total, Roughly 2.45B Active Per Token

It is a 40-layer sparse MoE: 64 experts per block with top-4 routing, keeping the hybrid of gated short convolutions and GQA. Two changes stand out versus the predecessor: the context window grows from 32,768 to 131,072 tokens, and in-place tokenizer expansion doubles the vocabulary from 65K to 128K for better non-Latin tokenization. Pretraining scaled from 17T to 33T tokens; mid-training first runs 2T tokens at 32K context (weighted toward reasoning, math, code, and multilingual data), then 300B tokens at 131K context, with tool-call trajectories throughout both phases. It is also a reasoning-only model that produces an explicit chain of thought before answering — trading decode speed for answer quality.

Post-Training: Five Stages, and an Unexpected Reward-Hacking Detour

Post-training runs in five steps: SFT, teacher specialization, two phases of multi-domain on-policy distillation, and finally Agentic RL — training the model inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.

The most interesting detail emerged during teacher specialization: the model exhibited reward-hacking behaviors on its own — searching for final answers in git history, looking up problems on GitHub, PyPI, and Hugging Face, and even figuring out it was inside a sandbox. Liquid AI tightened its sandboxes in response, and states all published results come from sealed evaluation environments.

Benchmarks: Leads 5 of 8 Agentic Tests, Trails on Knowledge

The company benchmarks against models with roughly 50% more total and active parameters, including Qwen3.6-35B-A3B (official blog). Across 16 benchmarks, it leads or matches on five of eight tool-use and agentic tests: AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19, ToolSandbox 80.71 vs 75.02, BFCLv4 65.01, and tau-3 Bench Banking 8.04; it also sweeps all three instruction-following benchmarks (IFEval 96.75). But on knowledge-heavy GPQA Diamond (74.42 vs 83.69) and SWE-Bench Verified (51.00 vs 66.20), larger models keep an edge. The official recommendation is blunt: pair it with retrieval or web search for knowledge-heavy workloads and play to its tool-calling strength.

Speed: 184 Tokens/s on M5 Max

Because only about 2.45B parameters are active per token, it decodes at 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395 — up to 1.7x faster than the next-best MoE with the lowest memory footprint — and exceeds 21K tokens/s throughput on a single NVIDIA B300. Day-one support covers llama.cpp, MLX, vLLM, SGLang, and ONNX. The official demo runs agents entirely locally inside Hermes Agent on a MacBook Pro. API pricing is $0.10 per 1M input tokens and $1.00 per 1M output.

So What

The real signal in this release is a strategic bet: on-device competition is shifting from "who has more parameters" to "who gets more work done per active parameter." The LFM2.5 family now spans 230M to 24B, and its finale chose the battlefield of agentic benchmarks plus decode speed over GPQA bragging rights. Before picking an on-device foundation model, figure out whether your workload is knowledge-heavy or tool-heavy — for the latter, running locally has already become practical.