LiquidAI Releases LFM2.5-24B-A2B Reasoning Model for On-Device AI Agents

LiquidAI has released LFM2.5-24B-A2B, a reasoning-focused open-weight model designed to power fast, private AI agents entirely on-device. The model uses a mixture-of-experts architecture with 24B total parameters but activates only 2B per token, making continuous agentic workloads affordable on consumer hardware.

Architecture and Training

LFM2.5-24B-A2B builds on the LFM2-24B-A2B architecture, keeping 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid design of gated short convolutions and grouped query attention. The context window grows from 32K to 131K tokens, and an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens for improved non-Latin language support. Unlike its predecessor, this is a reasoning-only model that produces an explicit chain of thought before answering.

Pre-training scaled from roughly 17T to about 33T tokens. Mid-training added 2T tokens at 32K context weighted toward reasoning, math, code, instruction following, and multilingual text, plus 300B tokens at 131K context weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases.

Post-training converts the base model into an agent in five stages: supervised fine-tuning, teacher specialization, multi-domain on-policy distillation for non-agentic tasks, a second distillation stage for agentic tasks, and agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.

Benchmark Performance

LiquidAI evaluated the model on 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math, and coding, comparing it against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning. Despite being the smallest model in the group, LFM2.5-24B-A2B scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench against 8.72 for the next-best model.

The model leads or matches competitors on five of eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and tau3-Bench Banking) and on all three instruction-following benchmarks (IFEval, IFBench, Multi-IF). Larger models retain an edge on knowledge-heavy benchmarks such as GPQA Diamond, HLE, and AA-Omniscience, as well as coding benchmarks like SWE-Bench Verified and SciCode. LiquidAI recommends pairing the model with retrieval or web search for knowledge-intensive workloads.

Inference Speed and Availability

The model ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX. On edge hardware, it decodes at 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, up to 1.7x faster than the next-best model while using the least memory. On a single NVIDIA B300 GPU, it reaches over 21,000 output tokens per second at high concurrency.

Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight and available on Hugging Face. The model requires transformers 5.0.0 or later and can be served locally behind any OpenAI-compatible backend for use with agent harnesses including Hermes Agent, llama.cpp, vLLM, SGLang, and LM Studio.

Sources: Hugging Face Blog