LFM2.5-24B-A2B is presented as a 24B-parameter reasoning model that uses 2B active parameters per token, designed to run private, agentic workflows on local hardware. The authors state the model produces an explicit chain of thought before final answers and is intended to call tools and handle multi-step tasks inside popular agent harnesses.
Architecture and training
The model retains the predecessor’s architecture: 40 layers, 64 experts per Mixture-of-Experts (MoE) block with top-4 routing, and a hybrid of gated short convolutions and GQA. Context capacity rises from 32K to 131K tokens, and an expanded tokenizer doubles the vocabulary from 65K to 128K tokens to improve non-Latin tokenization.
Pretraining was increased 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) and 300B tokens at 131K context (weighted toward long-context data). Tool calls and multi-turn trajectories were included during both phases.
Post-training conversion into an agent is performed in five stages named by the authors: Supervised fine-tuning (SFT); Teacher specialization; Multi-domain on-policy distillation (MOPD) non-agentic; MOPD agentic; and Agentic reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.
Benchmarks and performance
Evaluations compared LFM2.5-24B-A2B to Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math, and coding. The authors report that despite being the smallest model in the group, LFM2.5-24B-A2B is competitive with MoEs that have roughly 50% more total and active parameters.
Reported highlights include a LiveClaw score of 48.09 for LFM2.5-24B-A2B versus 38.19 for Qwen3.6-35B-A3B. The authors note LFM2.5-24B-A2B leads or matches on five of eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking) and leads on three instruction-following benchmarks. The blog also observes larger models retain an edge on knowledge-heavy and some coding benchmarks, recommending pairing this model with retrieval or web search for knowledge-heavy workloads.
Inference and usage
Day-one inference ecosystem support is reported for llama.cpp, MLX, vLLM, SGLang, and ONNX. Edge inference speeds are given as 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395; GPU decoding is reported to reach over 21K output tokens/s at high concurrency on a single NVIDIA B300.
The authors provide usage instructions noting transformers>=5.0.0 is required. Example install and load commands shown by the authors:
pip install -U transformers
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
model_id = "LiquidAI/LFM2.5-24B-A2B"
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
dtype="bfloat16",
)All performance claims and benchmark results are those reported by the authors.
Original source: Hugging Face Blog