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

What it claims

LFM2.5-24B-A2B is an upgrade of LFM2-24B-A2B with the same architecture. Against its predecessor:

• Context grows from 32K to 131K tokens.
• Vocabulary grows from 65K to 128K tokens, via in-place tokenizer expansion aimed at non-Latin scripts.
• Pre-training data roughly doubles, from ~17T to ~33T tokens.
• It becomes reasoning-only. It always emits a chain of thought before answering, so every request pays for reasoning tokens. Only 2B parameters are active per token, which keeps each of those tokens cheap.

Liquid AI pitches it as an on-device agent model. It is strongest at tool use and instruction following, and the page explicitly says it is weaker at knowledge and coding.

How it works

Mid-training has two phases, and tool calls and multi-turn trajectories appear in both:

1. 2T tokens at 32K context, weighted toward reasoning, math, code, instruction following and multilingual text.
2. 300B tokens at 131K context, weighted toward long-context data.

Post-training has five stages:

1. SFT.
2. Teacher specialization.
3. Multi-domain on-policy distillation (MOPD) from the specialist teachers on non-agentic tasks.
4. A second MOPD stage on agentic tasks.
5. Agentic RL run inside real harnesses: OpenClaw, Hermes Agent, Pi and OpenCode.

The page does not describe the teachers' sizes or the RL reward.

Evidence

Comparison set. The model is compared with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on 16 benchmarks. All were run under vLLM with each benchmark's recommended generation parameters. It is the smallest model in the group.

Tool use and agentic tasks. It leads or matches on 5 of 8 benchmarks: AutomationBench, BFCLv4, ToolSandbox, LiveClaw and τ³-Bench Banking.

• On LiveClaw it scores 48.09 against 38.19 for Qwen3.6-35B-A3B.
• On AutomationBench it scores more than 3× the next-best model.
• The page does not name the other three agentic benchmarks, where it trails.

Instruction following. It leads on all three benchmarks in this category.

Where it loses. Larger models do better on:

• Knowledge: GPQA Diamond, HLE and AA-Omniscience.
• Coding: SWE-Bench Verified (OpenCode) and SciCode.

Speed.

• 184 tok/s decode on an Apple M5 Max and 114 tok/s on a Ryzen AI Max+ 395. That is up to 1.7× faster than the next-best compared MoE, with the lowest memory use of the group. The page does not give the quantization or batch settings.
• Over 21K output tok/s on a single B300 at high concurrency.

How to use it

Two checkpoints are open-weight on Hugging Face: LiquidAI/LFM2.5-24B-A2B (the instruct/reasoning model) and LFM2.5-24B-A2B-Base. Day-one support is listed for llama.cpp, MLX, vLLM, SGLang and ONNX. The transformers path needs transformers>=5.0.0:

pip install -U transformers

The sample uses bfloat16 with device_map="auto". It notes that attn_implementation="flash_attention_2" can be enabled on a compatible GPU. Its generation settings are:

do_sample=True, temperature=0.1, top_k=50, repetition_penalty=1.05

To use it in an agent, serve it behind any OpenAI-compatible endpoint and point the harness at it. The Hermes Agent configuration is:

hermes config set model.provider custom
hermes config set model.base_url http://localhost:8080/v1
hermes config set model.default LFM2.5-24B-A2B
hermes config set model.context_length 131072
hermes config set model.api_mode chat_completions
hermes config set agent.tool_use_enforcement true

Default ports by backend:

Configurations for OpenClaw and Pi are on the model card, not on this page. Fine-tuning tutorials are linked.

Limits

• Knowledge and coding. Larger models are stronger here. The authors recommend pairing this model with retrieval or web search for knowledge-heavy workloads.
• No non-reasoning mode. Every response pays the latency of a chain of thought.
• No licence stated. The page says "open-weight" but names no licence. Check the model card before commercial use.
• Missing details.
• The page gives no per-benchmark numbers other than LiveClaw.
• It gives no memory figures behind "least memory".
• It gives no quantization settings for the edge speed results.
• Vendor results. All results are self-reported by Liquid AI.