What makes a 24-billion-parameter model run like a 2-billion one on your laptop?

The architecture that keeps most weights asleep

LFM2.5-24B-A2B keeps the skeleton of its predecessor LFM2-24B-A2B. Forty layers. Sixty-four experts in each mixture-of-experts block. A top-4 router wakes only four experts per token. The hybrid backbone mixes gated short convolutions with grouped-query attention. Two numbers changed. The context window grew from 32K to 131K tokens. An in-place tokenizer expansion doubled the vocabulary from 65K to 128K tokens. That improves non-Latin languages.

Pre-training consumed roughly 33 trillion tokens. Almost double the 17 trillion used for the base model. Mid-training added 2 trillion tokens at 32K context. The mix weighted toward reasoning, math, code, instruction following, and multilingual text. Then 300 billion tokens at the full 131K context. That phase weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases.

Five post-training stages turn a base model into an agent

Supervised fine-tuning comes first. Teacher specialization follows. Two rounds of multi-domain on-policy distillation arrive next. One round on non-agentic tasks. A second round focused on agentic work. The final stage runs agentic reinforcement learning inside actual harnesses. OpenClaw and Hermes Agent. Pi and OpenCode. The model learns to produce an explicit chain of thought before the final answer. It also learns to call tools and maintain multi-turn trajectories.

Benchmarks where the small model wins and where it does not

Sixteen benchmarks cover tool use, agentic workflows, knowledge, instruction following, math, and coding. Competitors: Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning. LFM2.5-24B-A2B is the smallest in the group.

It leads or matches on five of eight tool-use and agentic benchmarks. AutomationBench: 27.69 versus 8.72 for the next best. More than three times higher. BFCLv4: 65.01 versus 65.00. Essentially tied. ToolSandbox: 80.71 versus 75.02. LiveClaw: 48.09 versus 38.19. τ³-Bench Banking: 8.04 versus 7.63. It also leads on all three instruction-following benchmarks. IFEval 96.75. IFBench 74.69. Multi-IF 84.38.

Larger models keep an edge on knowledge-heavy tests. GPQA Diamond: 74.42 versus 83.69. HLE: 7.40 versus 7.69. AA-Omniscience: -24.90 versus -24.35. On coding, SWE-Bench Verified reaches 51.00 versus 66.20. SciCode: 32.71 versus 37.35. The announcement suggests pairing the model with retrieval or web search for knowledge-heavy workloads.

Speed on the hardware you already own

Day-one support spans llama.cpp, MLX, vLLM, SGLang, and ONNX. On an Apple M5 Max the model decodes 184 tokens per second. On a Ryzen AI Max+ 395 it reaches 114 tokens per second. That is up to 1.7 times faster than the next-best mixture-of-experts model while using the least memory. On a single NVIDIA B300 GPU it pushes past 21K output tokens per second at high concurrency.

Running it locally takes three commands

Install the latest transformers. Compatible with transformers>=5.0.0. Load the model with device_map="auto" and dtype="bfloat16". Generate with a chat template, temperature 0.1, top_k 50, repetition_penalty 1.05. To plug it into an agent harness, serve it behind any OpenAI-compatible backend. Hermes Agent example: set provider to custom, base URL to the local endpoint, default model to LFM2.5-24B-A2B, context length to 131072, API mode to chat completions, and tool use enforcement to true. Ports differ by backend. llama.cpp and MLX use 8080. vLLM uses 8000. SGLang uses 30000. LM Studio uses 1234.

Both the instruct and base weights are open and on Hugging Face today. Fine-tuning tutorials are linked from the model card.