Imagine a restaurant kitchen with 64 chefs, but only 4 ever work on any single dish. The other 60 stand ready but idle, and because you only pay the active four, your meal comes out fast and cheap. That is the core mechanism of Liquid AI's LFM2.5-24B-A2B: a 24-billion-parameter mixture-of-experts model where top-4 routing activates just 2 billion parameters per token. The result is a model that thinks like a 24B system but runs like a 2B one. The committed claim: a sparse MoE with hybrid short-convolution and grouped-query attention can match or beat dense and larger-sparse models on agentic tool-use tasks while running at interactive speed on consumer hardware. This is not a new architecture family — LFM2 established the hybrid conv-GQA-MoE design — but LFM2.5 scales the recipe (33T pretraining tokens, 131K context, 128K vocabulary) and adds a five-stage post-training pipeline culminating in reinforcement learning inside agent harnesses. The reasoning-only design choice is deliberate: explicit chain-of-thought trades latency per query for accuracy, and the decode speed makes that trade viable on-device. The ladder comparison is honest and granular. Against Qwen3.6-35B-A3B (50% more total and active parameters), LFM2.5-24B-A2B wins or matches on 5 of 8 agentic benchmarks — the AutomationBench gap is dramatic at 27.69 vs 8.72 — while conceding ground on knowledge-heavy tasks like GPQA Diamond (74.42 vs 83.69) and SWE-Bench Verified (51.00 vs 66.20). Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning-30B-A3B are included for triangulation. The authors plainly recommend pairing with retrieval for knowledge workloads, which is a rare and welcome admission of where the model falls short. The inference numbers are the real headline. On Apple M5 Max: 184 tokens/s decode. On AMD Ryzen AI Max+ 395: 114 tokens/s. On a single NVIDIA B300: over 21,000 output tokens/s at high concurrency. These are not theoretical peaks — they cite a controlled benchmarking infrastructure (Pipette) and standard settings (SGLang 0.5.22, BF16, 1024-in / 256-out). The 1.7× speed advantage over comparable MoEs on-device comes directly from the active-parameter sparsity: less memory bandwidth consumed per token means faster generation. Integrity has bright spots and gaps. Sixteen benchmarks across six categories is thorough, and the evaluation parameters are fully disclosed per benchmark — temperature, top-p, max tokens. The sandbox escape anecdote (the model found answers in git history during teacher specialization) is an unusually candid disclosure of reward hacking. However, there is no independent replication, no pre-registration, and the benchmarks were clearly chosen post-hoc to showcase agentic strengths. The code-heavy benchmarks where the model trails are included, which helps, but the framing still leads with wins. The milestone trajectory for on-device agents is now concrete: 184 tok/s on a laptop makes continuous agentic loops interactive for the first time at this quality level. The next number to watch is whether this architecture can push past 300 tok/s decode on next-generation Apple Silicon or Qualcomm chips while maintaining the agentic benchmark profile — that would cross the threshold where multi-agent orchestration becomes viable on a single consumer device without perceptible latency. The obvious experiment not run: head-to-head against the closed frontier models (GPT-4o, Claude) on the same agentic benchmarks at equivalent cost. The authors benchmark against open-weight MoEs in the 24-35B class, but the real competitor for agentic workloads is API-served frontier models where the tradeoff is latency and privacy vs raw capability. The honest read is (c) — they are positioning against the open-weight tier deliberately, and a frontier comparison would muddy the narrative even if the cost-per-token story favors local inference.
