Run fast and private on-device agents with LFM2.5-24B-A2B.

Meta description: LFM2.5-24B-A2B is a local reasoning agent model with 2 billion active parameters, 184 tokens per second on Apple M5 Max, and a 131,000-token context window.

Tags: LFM2.5-24B-A2B, reasoning agents, local inference, Mixture-of-Experts, open weights

Featured image: abstract friendly illustration of a small chip-shaped network activating four expert nodes inside a larger model body.

LFM2.5-24B-A2B is a reasoning-only agent model designed for developers who want AI applications to think before answering, call tools, and run on local hardware rather than relying on a much larger model. It is a Mixture-of-Experts model, meaning it has many parameters but activates only a subset for each token; here that active count is 2 billion parameters per token. The release says the model scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, a model with about 50 percent more total and active parameters, making it competitive on agentic and tool-use tasks while staying efficient.

Why this release matters for agent builders

The value is not just a new benchmark score. Agent workflows often require a model to decide which tool to call, check the result, revise its answer, and keep doing that across several steps. That kind of work can be expensive if the model is large, slow, or needs a hosted service. LFM2.5-24B-A2B aims at that cost point: it always produces an explicit chain of thought before the final answer, but only 2 billion parameters are active per token. The practical effect, as described by the release, is that each reasoning token is cheap enough to support local or edge deployment.

This makes the model most interesting for teams building assistants that need reliable tool use rather than broad factual recall. If the task is to check a tool result, revise an answer, and continue through several steps, the benchmark profile suggests the model can compete with larger Mixture-of-Experts systems. If the task is to answer a specialized question from memory, the release itself recommends pairing it with retrieval or web search.

What LFM2.5-24B-A2B changes under the hood

LFM2.5-24B-A2B builds on LFM2-24B-A2B and keeps the same broad architecture: 40 layers, 64 experts per Mixture-of-Experts block, and top-4 routing. That routing means four experts are used for each token. It also keeps a hybrid of gated short convolutions and grouped-query attention. Grouped-query attention is a common technique for reducing memory and compute during inference by sharing key-value information across query groups.

The context window grows from 32,000 to 131,000 tokens, which matters for long documents and multi-step tool histories. The tokenizer is expanded in place, doubling vocabulary from 65,000 to 128,000 tokens, with better non-Latin tokenization. That update can improve how the model handles multilingual text without replacing the tokenizer from scratch.

Training also expanded. Pre-training scaled from about 17 trillion tokens to about 33 trillion tokens. Mid-training then adds 2 trillion tokens at 32,000 context, weighted toward reasoning, math, code, instruction following, and multilingual text, plus 300 billion tokens at 131,000 context, weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases.

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

How it compares on reasoning and tool use

The release evaluated LFM2.5-24B-A2B against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning on 16 benchmarks covering tool use and agentic workflows, knowledge, instruction following, math, and coding. All models were evaluated with vLLM using each benchmark's recommended generation parameters. LFM2.5-24B-A2B is the smallest model in that group, yet it is competitive with Mixture-of-Experts systems that have about 50 percent more total and active parameters.

The clearest strengths are tool use and instruction following. It leads or matches the competition on five of the eight tool-use and agentic benchmarks: AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. On AutomationBench, it scores more than three times the next-best model. It also leads on all three instruction-following benchmarks.

The larger models still keep advantages in some areas. Knowledge-heavy benchmarks such as GPQA Diamond, HLE, and AA-Omniscience, plus coding benchmarks such as SWE-Bench Verified (OpenCode) and SciCode, favor the bigger systems. That split suggests a practical pattern: use LFM2.5-24B-A2B as the efficient agent layer, then bring in retrieval, web search, or a larger model when the task needs deep factual recall or heavy code generation.

Running the model locally and inside agent stacks

LFM2.5-24B-A2B ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX. On edge hardware, the release reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395. That is up to 1.7 times faster than the next-best compared model, while using the least memory. On a single NVIDIA B300 GPU, it reaches over 21,000 output tokens per second at high concurrency.

For local use, the model requires a transformers version compatible with transformers>=5.0.0. The release also describes serving it locally behind an OpenAI-compatible backend and pointing an agent harness at that endpoint. The port depends on the backend: llama.cpp and MLX use 8080, vLLM uses 8000, SGLang uses 30000, and LM Studio uses 1234. Model card configurations are available for OpenClaw and Pi, and the release blog includes video examples.

Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight and available on Hugging Face. The release also links fine-tuning tutorials for adapting the model to specific tasks. If the work relies on the model's results, the release blog is the citation source.

Who should use it, and where it may fall short

LFM2.5-24B-A2B is a strong fit for developers who need a local reasoning agent with tool calling, long context, and fast token generation on edge hardware. It is especially useful when the application spends many tokens reasoning, checking tools, and continuing through several steps, because the active-parameter design keeps that work cheap.

It is less suitable for projects that need a single model to answer broad knowledge questions or handle coding-heavy workloads without retrieval. In those cases, the release's own guidance is clear: pair the model with search, retrieval, or a larger system for knowledge-heavy and coding-heavy workloads. The practical takeaway is that LFM2.5-24B-A2B looks most valuable as an efficient agent layer, not as a replacement for every larger model.