LFM2.5-24B-A2B: Built for Fast On-Device Agents | Blog

Today, we launch LFM2.5-24B-A2B, a 24B-parameter reasoning mannequin with simply 2B lively parameters, constructed for quick, non-public brokers on-device. It is the most important and closing mannequin within the LFM2.5 household, which now spans two orders of magnitude: from LFM2.5-230M to LFM2.5-24B-A2B.

LFM2.5-24B-A2B is aggressive with Qwen3.6-35B-A3B, a mannequin with ~50% extra whole and lively parameters, on 5 of eight agentic and tool-use benchmarks, scoring 48.09 on LiveClaw versus 38.19 and 27.69 on AutomationBench versus 8.72.  Because solely 2B parameters are lively per token, it runs as much as 1.7× quicker than comparable MoEs on-device, with 184 tokens/s on an Apple M5 Max and as much as 21K output tokens/s on a single NVIDIA B300 GPU.

The base (LFM2.5-24B-A2B-Base) and post-trained (LFM2.5-24B-A2B) fashions are open-weight and accessible right this moment on Hugging Face and our Playground. Check out our docs on find out how to run and fine-tune it domestically.

What modified since LFM2-24B-A2B

LFM2.5-24B-A2B builds on our earlier LFM2-24B-A2B launch, with an expanded vocabulary and context window, scaled-up pretraining, and large-scale reinforcement studying. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only mannequin that produces an express chain of thought earlier than its closing reply. We adopted this technique to leverage its excessive decode pace and enhance reply high quality.

Architecture. It retains the 40 layers, 64 specialists per MoE block with top-4 routing, and the identical hybrid of gated quick convolutions and GQA. The two architectural modifications are the context window extension from 32,768 to 131,072 tokens, and the in-place tokenizer expansion, which doubles the vocabulary from 65K to 128K tokens for higher non-Latin tokenization.

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

Mid-training has two phases. The first runs 2T tokens at 32K context on curated, higher-quality information weighted towards reasoning, math, code, instruction following, and multilingual textual content. The second runs 300B tokens at 131K context, weighted towards lengthy context information. Tool calls and multi-turn trajectories seem all through each phases to arrange the mannequin for agentic workloads.

During instructor specialization, LFM2.5-24B-A2B exhibited well-known reward-hacking behaviors: looking for the ultimate reply in a repository’s git historical past or on platforms like GitHub, PyPI, and Hugging Face. More surprisingly, it even discovered it was in a sandbox, as proven within the examples beneath.

This habits seems to be emergent however is likely to be defined by coaching information containing related patterns. We tightened our sandboxes in response, and all reported outcomes beneath come from sealed analysis environments.

Benchmarks

We evaluated LFM2.5-24B-A2B throughout 16 benchmarks overlaying software use and agentic workflows, information, instruction following, math, and coding. Despite being the smallest mannequin in whole and lively variety of parameters, it’s aggressive with, and infrequently outperforms, bigger MoEs with ~50% extra whole and lively parameters, similar to Qwen3.6-35B-A3B.

LFM2.5-24B-A2B leads or matches the competitors on 5 of the eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking), with its widest margin on AutomationBench, the place it scores greater than 3 times the next-best mannequin. It additionally leads on all three instruction-following benchmarks. Larger fashions maintain an edge on knowledge-heavy benchmarks like GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks like SWE-Bench Verified (OpenCode) and SciCode.

The outcomes make LFM2.5-24B-A2B a really perfect basis for brokers that have to purpose, name instruments, and run reliably in your machine. For knowledge-heavy workloads, we advocate pairing LFM2.5-24B-A2B with retrieval or net search, which performs to its tool-calling power.

Fast inference in every single place

LFM2.5-24B-A2B ships with day-one assist throughout the inference ecosystem:

• llama.cpp — GGUF checkpoints for environment friendly edge inference
• MLX — Optimized inference for Apple Silicon
• vLLM — GPU-accelerated serving for manufacturing throughput
• SGLang — High-performance GPU serving framework
• ONNX — Cross-platform inference throughout numerous accelerators

All runs had been collected on Pipette, the identical benchmarking roads behind Liquid AI’s public machine efficiency information.

Edge inference. Due to the environment friendly LFM2 structure, LFM2.5-24B-A2B is the quickest of the in contrast MoEs at studying in prompts and producing solutions. It decodes 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, as much as 1.7x quicker than the next-best mannequin, whereas utilizing the least reminiscence.

GPU inference. To reveal the competitiveness of LFM structure we evaluate the GPU inference throughput to different equally sized MoE fashions. We measure GPU throughput (whole output tokens / wall time) on a single NVIDIA B300 GPU utilizing a sustained-load setting: at every concurrency degree, we repeatedly keep the goal variety of in-flight requests, changing every accomplished request instantly.

We benchmark every mannequin with SGLang 0.5.22, 1,024 enter tokens, as much as 256 output tokens, in BF16. LFM2.5-24B-A2B is the quickest mannequin in its measurement class, reaching over 21K output tokens per second at excessive concurrency.

Run native brokers with LFM2.5-24B-A2B

LFM2.5-24B-A2B’s measurement, pace, and capabilities make it a fantastic selection for brokers that run repeatedly by yourself {hardware}. In the demo beneath, we run it inside Hermes Agent on an M5 Max MacE book Pro, the place it plans, calls instruments, and works by an actual process fully on-device, with none cloud API calls.

The level of the demo shouldn’t be the person instruments. It is that the tool-dispatch loop feels interactive on client {hardware}: ask, suggest, affirm, run, repeat, all in properly underneath a second per dispatch, with full audit trails and your information by no means leaving the machine. Check out our local agent guide for find out how to serve the mannequin domestically and join it with the agent harness of your selection.

Get Started

Start constructing right this moment with LFM2.5-24B-A2B, accessible on Hugging Face.

For API inference, it’s also possible to join Liquid Console the place you will obtain $2 in credit score to get began. The mannequin is priced at $.10/1M enter tokens (at $.05 for cached enter) and $1.00/1M output tokens.

With LFM2.5, we’re delivering on our imaginative and prescient of AI that runs anyplace. These fashions are:

• Open-weight — Download, fine-tune, and deploy. See our model license.
• Fast from day one — Native assist for llama.cpp, MLX, and vLLM throughout Apple, AMD, Qualcomm, and Nvidia {hardware}
• An entire household — From base fashions for personalization to specialised audio and imaginative and prescient variants, one structure covers numerous use instances

We cannot wait to see what you construct.

Download on Hugging FaceRun your harness of choice locally with LFM2.5-24B-A2BIntegrate with our API

Citation

For citations, please use the next reference or BibTeX:

@article{liquidAI202624BA2B,
  writer = {Liquid AI},
  title = {LFM2.5-24B-A2B: Built for Fast On-Device Agents},
  journal = {Liquid AI Blog},
  yr = {2026},
  word = {www.liquid.ai/weblog/lfm2-5-24b-a2b},
}

_{*All the fashions had been evaluated with vLLM and the next technology parameters:}

• _{AutomationBench: temperature = 0.6, top_p = 0.95, no max output tokens (bounded by the mannequin’s context window).}
• _{BFCLv4: temperature = 0.001, max output tokens = 4096.}
• _{ToolSandbox: temperature = 0, max output tokens = 1024; person simulator: GPT-4o-2024-08-06, temperature = 0, max output tokens = 1024.}
• _{LiveClaw: temperature = 1.0 (vLLM default; the agent units no temperature), max output tokens = 8192.}
• _{τ³-Bench Banking: official person mannequin (GPT-5.2, low reasoning effort) with default BM25 retrieval, temperature = 0, max output tokens = 16384 per flip.}
• _{WebArena: temperature = 0.6, top_p=0.95, max_turns=40 (no of browser actions allowed), no max output tokens (bounded by the mannequin’s context window))}
• _{Other evals: temperature = 0.6, max output tokens = 32768.}