A 24B reasoning model that activates only 2B parameters per token leads or matches on five of eight tool-use benchmarks

On-device agents no longer have to depend on a cloud model: a reasoning model with 24B total parameters and only 2B active per token can run continuously on local hardware and call tools. Comparable MoE models were either larger or lacked the think-before-answering behavior; LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer. It was evaluated against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math and coding, and is the smallest model in that group.

The model keeps the predecessor architecture and turns the base into an agent through expanded pre-training and mid-training data plus a five-stage post-training pipeline. The architecture stays at 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and GQA; what changes is the context window, from 32K to 131K tokens, and an in-place tokenizer expansion that doubles the vocabulary from 65K to 128K. Pre-training scaled from about 17T to about 33T tokens, mid-training added 2T tokens at 32K context and 300B tokens at 131K context, with tool calls and multi-turn trajectories in both phases; post-training runs supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation, agentic multi-domain on-policy distillation, and agentic reinforcement learning.

It leads or matches larger models on tool use and instruction following, while knowledge-heavy and coding tasks still favor the larger models. It leads or matches on five of eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw and τ³-Bench Banking, and leads on all three instruction-following benchmarks; larger models keep an edge on GPQA Diamond, HLE, AA-Omniscience, SWE-Bench Verified and SciCode. On AutomationBench its 27.69 is more than three times the next-best 8.72, and LiveClaw is 48.09 versus 38.19; for speed it decodes 184 tokens/s on an M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, up to 1.7x faster than the next-best model, and exceeds 21K output tokens/s at high concurrency on a single B300.

This report covers the LFM2.5-24B-A2B release: model specifications, the training pipeline, results on 16 benchmarks, inference speed, and open-weight availability with day-one inference-ecosystem support. The intended reader is an engineer who needs to run reasoning and tool-calling agents on their own hardware. The material recommends pairing the model with retrieval or web search for knowledge-heavy workloads and gives configuration steps for pointing an agent harness at a local OpenAI-compatible backend.

The material states that larger models keep an edge on knowledge-heavy benchmarks such as GPQA Diamond, HLE and AA-Omniscience, and on coding benchmarks such as SWE-Bench Verified and SciCode, so it recommends pairing the model with retrieval or web search for knowledge-heavy workloads. All models were evaluated with vLLM using each benchmark's recommended generation parameters, and the speed figures come from specific M5 Max, Ryzen AI Max+ 395 and single-B300 configurations.