Glonce

Liquid AI releases LFM2.5-24B-A2B, an open-weight 2B-active reasoning agent model

modelsThe Glonce editorsHugging Face🔥64Glonce rating64Impact55Novelty55Relevance100Surprise35Players40Credibility80
Liquid AI releases LFM2.5-24B-A2B, an open-weight 2B-active reasoning agent model

Liquid AI has released LFM2.5-24B-A2B, a mixture-of-experts model built for agents and tool use that runs on local hardware. Only 2B parameters are active per token. Both the model and its Base variant, LFM2.5-24B-A2B-Base, are open-weight and available on Hugging Face.

The model builds on LFM2-24B-A2B and keeps its architecture: 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and GQA. The context window grows from 32K to 131K tokens, and an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens for better non-Latin tokenization. Unlike its predecessor, it is a reasoning-only model: it always produces an explicit chain of thought before its final answer, and with 2B active parameters each reasoning token is cheap.

Training was scaled up. Pre-training grew from about 17T to about 33T tokens. Mid-training then added 2T tokens at 32K context, weighted toward reasoning, math, code, instruction following and multilingual text, and 300B tokens at 131K context, weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases. Post-training has five stages: supervised fine-tuning; teacher specialization; multi-domain on-policy distillation (MOPD) on non-agentic tasks; a second MOPD stage focused on agentic tasks; and agentic reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

Liquid AI compared the model with 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 run with vLLM using each benchmark's recommended generation parameters. The company calls LFM2.5-24B-A2B the smallest model in the group, yet competitive with MoEs that have about 50% more total and active parameters. The headline figure is LiveClaw: 48.09 against 38.19 for Qwen3.6-35B-A3B. It leads or matches the competition on five of the eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw and τ³-Bench Banking), scores more than three times the next-best model on AutomationBench, and leads on all three instruction-following benchmarks. Larger models keep an edge on knowledge-heavy benchmarks (GPQA Diamond, HLE, AA-Omniscience) and on coding benchmarks (SWE-Bench Verified with OpenCode, and SciCode), so Liquid AI advises pairing the model with retrieval or web search for knowledge-heavy work.

On speed, Liquid AI reports 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395. It describes that as the fastest of the compared MoEs on edge hardware, up to 1.7x faster than the next-best model, while using the least memory. On a single NVIDIA B300 GPU at high concurrency it reaches over 21K output tokens per second, which the company calls the fastest in its size class. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX.

The post also shows how to run it: install transformers 5.0.0 or later, load LiquidAI/LFM2.5-24B-A2B in bfloat16, or serve it behind any OpenAI-compatible backend and point an agent harness such as Hermes Agent at it. The model card covers OpenClaw and Pi configs.

Key facts

  • LFM2.5-24B-A2B is a reasoning-only MoE with 2B active parameters per token, released as open weights together with the Base variant on Hugging Face.
  • Liquid AI reports 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, a model with about 50% more total and active parameters; it leads or matches on five of eight tool-use and agentic benchmarks.
  • Reported speeds: 184 tok/s on an Apple M5 Max, 114 tokens/s on a Ryzen AI Max+ 395 (up to 1.7x the next-best model), and over 21K output tok/s on a single NVIDIA B300 at high concurrency.
  • Context grows from 32K to 131K tokens, vocabulary doubles from 65K to 128K, and pre-training rose from about 17T to about 33T tokens.
  • Larger models keep an edge on knowledge-heavy benchmarks (GPQA Diamond, HLE, AA-Omniscience) and coding benchmarks (SWE-Bench Verified with OpenCode, SciCode).

Why it matters

Agent models usually need a lot of compute, because every reasoning step costs tokens. Liquid AI is betting that a mixture-of-experts design with only 2B active parameters out of 24B can reason before answering and still stay cheap per token. The company says it holds up against Qwen3.6-35B-A3B on agentic and tool-use tasks (48.09 vs 38.19 on LiveClaw) while being the smallest model in its comparison group. It also reports fast decoding on laptop-class chips, which is the point of the on-device pitch.

Who it affects

Developers building agents that call tools and need to run locally on laptops, workstations or their own GPUs. Teams already using harnesses such as OpenClaw, Hermes Agent or Pi can point them at the model through an OpenAI-compatible server. Researchers and fine-tuners get the Base variant as well. People who need strong knowledge recall or coding help are not the target: the blog itself says larger models lead there.

How to use it

Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are on Hugging Face as open weights. For a direct run, install transformers (compatible with 5.0.0 or later), load LiquidAI/LFM2.5-24B-A2B with AutoModelForCausalLM in bfloat16 with device_map set to auto, and generate with the chat template. The example uses temperature 0.1, top_k 50 and repetition_penalty 1.05. For agents, serve it behind any OpenAI-compatible backend and configure the harness; the post gives Hermes Agent settings with a context length of 131072. Default ports depend on the backend: llama.cpp and MLX use 8080, vLLM 8000, SGLang 30000 and LM Studio 1234. Liquid AI also links fine-tuning tutorials and a guide for running a local agent. The text names no license for the open weights.

How solid is it

This is a first-party announcement. The benchmarks were run by Liquid AI, and no independent evaluation is mentioned. The evaluation covers 16 benchmarks against three named models, with all models run in vLLM under each benchmark's recommended generation parameters. The text gives only a few scores directly: LiveClaw (48.09 vs 38.19) and a more-than-three-times ratio on AutomationBench. The full benchmark table is not in the text. The source also does not say which model is the 'next-best' for the 1.7x speed claim, nor the precision or quantization used in the speed measurements.

Risks and caveats

The model leads or matches on only five of the eight tool-use and agentic benchmarks, and the text does not say which three it misses. Larger models keep an edge on GPQA Diamond, HLE and AA-Omniscience, and on SWE-Bench Verified (OpenCode) and SciCode, so Liquid AI recommends retrieval or web search for knowledge-heavy workloads. Being reasoning-only, it always produces a chain of thought before the answer. The memory claim is stated as 'the least memory' without figures, and the speed numbers are Liquid AI's own measurements on specific hardware.

“It is the smallest model in the group, yet it is competitive with MoEs with ~50% more total and active parameters.”

— Liquid AI, release blog for LFM2.5-24B-A2B