LiquidAI Releases LFM2.5-24B-A2B for Fast On-Device Agents

LiquidAI LFM2.5-24B-A2B release

LiquidAI just dropped LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts model that only wakes up 2 billion parameters per token. That’s the headline: you can actually run serious reasoning agents on local hardware now. It landed on Hugging Face October 13, 2026, complete with benchmarks and integration guides for the usual agent frameworks. Since each reasoning step touches a tiny slice of the total weights, the model chews through long chains of thought and tool-calling loops without melting your consumer CPU or GPU.

LFM2.5-24B-A2B architecture specs

The skeleton is familiar-40 layers, 64 experts, top-4 routing, just like the LFM2-24B-A2B predecessor-but they stretched the context window from 32K to 131K tokens and doubled the vocabulary to 128K for way better non-Latin support. Pre-training data ballooned from roughly 17 trillion to 33 trillion tokens. A mid-training phase tossed in 2 trillion tokens at 32K context, heavily weighted toward reasoning, math, code, instruction following, and multilingual data, plus 300 billion tokens at the full 131K length for long-sequence chops. Tool calls and multi-turn trajectories were baked in throughout both phases. Post-training ran five laps: supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation covering non-agentic and agentic tasks, and a final agentic RL pass inside harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode.

LFM2.5-24B-A2B benchmark results

Benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning show this thing leading or matching bigger rivals on five of eight tool-use and agentic tests. On AutomationBench it hit 27.69-more than triple the next best score of 8.72. It also topped BFCLv4 at 65.01, ToolSandbox at 80.71, LiveClaw at 48.09, and τ³-Bench Banking at 8.04. Instruction-following suites IFEval, IFBench, and Multi-IF all leaned its way. Bigger models still win on knowledge-heavy stuff like GPQA Diamond and HLE, plus coding benchmarks such as SWE-Bench Verified and SciCode, so the LiquidAI crew suggests pairing this with retrieval or web search for those workloads.

Speed is the real pitch. On an Apple M5 Max it decodes at 184 tokens per second; on a Ryzen AI Max+ 395 it manages 114-both faster than the compared MoEs while sipping the least memory. On a single NVIDIA B300 GPU it can push past 21,000 output tokens per second at high concurrency. Day-one support covers llama.cpp, MLX, vLLM, SGLang, and ONNX, so you can drop it into existing pipelines without waiting for the ecosystem to catch up.

LFM2.5-24B-A2B installation requirements

Getting started needs transformers 5.0.0 or newer. The model ID is LiquidAI/LFM2.5-24B-A2B, and the recommended generation config uses a low 0.1 temperature with top-k 50 and a modest repetition penalty. For agent deployments, serve it locally behind any OpenAI-compatible backend-llama.cpp and MLX default to port 8080, vLLM to 8000, SGLang to 30000, LM Studio to 1234-then point harnesses like Hermes Agent, OpenClaw, or Pi at the base URL and model name in their config. The model card has example configs and the release blog links to video demos.

If you’re building private, on-device agents, treat this as a strong candidate for the reasoning layer, especially when tool use and instruction following matter more than raw knowledge retrieval. The combo of a 131K context window, explicit chain-of-thought output, and sub-2B active parameter count makes continuous agent loops on laptops and workstations genuinely feasible-no cloud required. Teams already on OpenAI-compatible serving stacks can swap it in with minimal plumbing, and the benchmark profile suggests it’ll outrun larger MoEs on the agentic tasks that usually bottleneck real-world workflows.

Source: Hugging Face