LFM2.5-24B-A2B: Built for fast on-device agents

Today, we publish LFM2.5-24B-A2B, a 24B-parameter reasoning example alongside fair 2B energetic parameters, built for fast, personal agents on-device. It is the largest and final example in the LFM2.5 family, which now spans two orders of magnitude: from LFM2.5-230M to LFM2.5-24B-A2B.

LFM2.5-24B-A2B is rivalrous alongside Qwen3.6-35B-A3B, a example alongside ~50% additional total and energetic parameters, on five of eight agentic and tool-use benchmarks, scoring 48.09 on LiveClaw versus 38.19 and 27.69 on AutomationBench versus 8.72.  Because lone 2B parameters are energetic per token, it runs up to 1.7× faster than comparable MoEs on-device, alongside 184 tokens/s on an Apple M5 Max and up to 21K output tokens/s on a sole NVIDIA B300 GPU.

The basis (LFM2.5-24B-A2B-Base) and post-trained (LFM2.5-24B-A2B) models are open-weight and accessible today on Hugging Face and our Playground. Check out our docs on how to run and fine-tune it locally.

What changed since LFM2-24B-A2B

LFM2.5-24B-A2B builds on our former LFM2-24B-A2B release, alongside an expanded vocabulary and environment window, scaled-up pretraining, and large-scale reinforcement learning. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only example that produces an definitive sequence of idea before its final answer. We adopted this scheme to leverage its elevated decode speed and enhance answer quality.

Architecture. It keeps the 40 layers, 64 experts per MoE obstacle alongside top-4 routing, and the identical hybrid of gated abbreviated convolutions and GQA. The two architectural changes are the environment opening expansion from 32,768 to 131,072 tokens, and the in-place tokenizer expansion, which doubles the vocabulary from 65K to 128K tokens for improved non-Latin tokenization.

Pre-training was scaled up from 17T tokens to 33T tokens.

Mid-training has two phases. The archetypal runs 2T tokens at 32K environment on curated, higher-quality data valued toward reasoning, math, code, education following, and multilingual text. The second runs 300B tokens at 131K context, valued toward lengthy environment data. Tool calls and multi-turn trajectories appear throughout the two phases to prepared the example for agentic workloads.

During instructor specialization, LFM2.5-24B-A2B exhibited well-known reward-hacking behaviors: searching for the final answer in a repository's git former or on platforms akin GitHub, PyPI, and Hugging Face. More surprisingly, it equal figured out it was in a sandbox, as shown in the examples below.

This behavior looks emergent but power be explained by training data containing akin patterns. We tightened our sandboxes in response, and all reported results below arrive from sealed evaluation environments.

Benchmarks

We evaluated LFM2.5-24B-A2B throughout 16 benchmarks covering tool use and agentic workflows, knowledge, education following, math, and coding. Despite being the smallest example in total and energetic figure of parameters, it is rivalrous with, and frequently outperforms, larger MoEs alongside ~50% additional total and energetic parameters, specified as Qwen3.6-35B-A3B.

LFM2.5-24B-A2B leads or matches the competition on five of the eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking), alongside its widest border on AutomationBench, anywhere it scores additional than three times the next-best model. It additionally leads on all three instruction-following benchmarks. Larger models keep an border on knowledge-heavy benchmarks akin GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks akin SWE-Bench Verified (OpenCode) and SciCode.

The results create LFM2.5-24B-A2B an ideal basis for agents that need to reason, call tools, and run reliably on your device. For knowledge-heavy workloads, we propose pairing LFM2.5-24B-A2B alongside retrieval or web search, which plays to its tool-calling strength.

Fast conclusion everywhere

LFM2.5-24B-A2B ships alongside day-one assistance throughout the conclusion ecosystem:

• llama.cpp — GGUF checkpoints for productive border inference
• MLX — Optimized conclusion for Apple Silicon
• vLLM — GPU-accelerated serving for manufacturing throughput
• SGLang — High-performance GPU serving framework
• ONNX — Cross-platform conclusion throughout varied accelerators

All runs were collected on Pipette, the identical benchmarking infrastructure rearward Liquid AI's community equipment achievement data.

Edge inference. Due to the productive LFM2 architecture, LFM2.5-24B-A2B is the fastest of the compared MoEs at study in prompts and generating answers. It decodes 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, up to 1.7x faster than the next-best model, during using the smallest memory.

GPU inference. To display the competitiveness of LFM architecture we difference the GPU conclusion throughput to another likewise sized MoE models. We measure GPU throughput (total output tokens / partition time) on a sole NVIDIA B300 GPU using a sustained-load setting: at all concurrency level, we continuously keep the mark figure of in-flight requests, replacing all completed petition immediately.

We benchmark all example alongside SGLang 0.5.22, 1,024 input tokens, up to 256 output tokens, in BF16. LFM2.5-24B-A2B is the fastest example in its size class, reaching complete 21K output tokens per second at elevated concurrency.

Run local agents alongside LFM2.5-24B-A2B

LFM2.5-24B-A2B’s size, speed, and capabilities create it a awesome choice for agents that run continuously on your own hardware. In the demo below, we run it inner Hermes Agent on an M5 Max MacBook Pro, anywhere it plans, calls tools, and plant through a genuine project entirely on-device, without any haze API calls.

The item of the demo is not the idiosyncratic tools. It is that the tool-dispatch iteration feels interactive on person hardware: ask, propose, confirm, run, repeat, all in fine under a second per dispatch, alongside complete audit trails and your data never leaving the device. Check out our local delegate guide for how to assist the example locally and nexus it alongside the delegate harness of your choice.

Get Started

Start construction today alongside LFM2.5-24B-A2B, accessible on Hugging Face.

For API inference, you can additionally sign up for Liquid Console anywhere you'll obtain $2 in credit to get started. The example is priced at $.10/1M input tokens (at $.05 for cached input) and $1.00/1M output tokens.

With LFM2.5, we're delivering on our imagination of AI that runs anywhere. These models are:

• Open-weight — Download, fine-tune, and deploy. See our model license.
• Fast from day one — Native assistance for llama.cpp, MLX, and vLLM throughout Apple, AMD, Qualcomm, and Nvidia hardware
• A complete family — From basis models for customization to specialized audio and imagination variants, one architecture covers varied use cases

We can't delay to see what you build.

Download on Hugging Face

Run your harness of choice locally alongside LFM2.5-24B-A2B

Integrate alongside our API

Citation

For citations, delight use the following citation or BibTeX:

_{*All the models were evaluated alongside vLLM and the following generation parameters:}

• _{AutomationBench: heat = 0.6, top_p = 0.95, no max output tokens (bounded by the model's environment window).}
• _{BFCLv4: heat = 0.001, max output tokens = 4096.}
• _{ToolSandbox: heat = 0, max output tokens = 1024; person simulator: GPT-4o-2024-08-06, heat = 0, max output tokens = 1024.}
• _{LiveClaw: heat = 1.0 (vLLM default; the delegate sets no temperature), max output tokens = 8192.}
• _{τ³-Bench Banking: authoritative person example (GPT-5.2, low reasoning effort) alongside default BM25 retrieval, heat = 0, max output tokens = 16384 per turn.}
• _{WebArena: heat = 0.6, top_p=0.95, max_turns=40 (no of browser actions allowed), no max output tokens (bounded by the model's environment window))}
• _{Other evals: heat = 0.6, max output tokens = 32768.}