Liquid AI releases LFM2.5-24B-A2B, an open 24B MoE with 2B active parameters
Serguey Asael Shinder2 min readLiquid AI published LFM2.5-24B-A2B on the Hugging Face blog on 13 October 2026: a mixture-of-experts language model with 24 billion parameters in total and about 2 billion active per token, released as open weights together with a base model, LFM2.5-24B-A2B-Base.
What changed against LFM2-24B-A2B. The architecture is kept: 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and grouped-query attention. The context window grows from 32K to 131K tokens, and an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens, which the authors tie to better tokenization of non-Latin scripts. The bigger behavioural change is that the new model is reasoning-only: it always produces an explicit chain of thought before answering.
How it was trained. Pre-training was scaled from roughly 17 trillion to 33 trillion tokens. Mid-training added 2T tokens at 32K context, weighted towards reasoning, maths, code, instruction following and multilingual text, plus 300B tokens at 131K context for long-context data. Post-training runs in five stages, including two rounds of multi-domain on-policy distillation and agentic reinforcement learning inside agent harnesses the post names — OpenClaw, Hermes Agent, Pi and OpenCode.
What the numbers say, and who chose them. Liquid AI compares the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on 16 benchmarks. By its account the model leads or matches on five of eight tool-use and agentic benchmarks — on LiveClaw it reports 48.09 against 38.19 for Qwen — and on all three instruction-following benchmarks. The same post states the limits: larger models keep the edge on knowledge-heavy tests such as GPQA Diamond, HLE and AA-Omniscience and on coding benchmarks such as SWE-Bench Verified and SciCode, and the authors suggest pairing the model with retrieval for knowledge-heavy work. All results are the vendor's, run with vLLM.
Speed and deployment. The reported decoding speed is 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, and over 21,000 output tokens/s at high concurrency on a single NVIDIA B300. The model ships with support in llama.cpp, MLX, vLLM, SGLang and ONNX and requires transformers 5.0 or later; the post shows it served behind an OpenAI-compatible endpoint for local agents.
For teams evaluating on-device agents, the useful takeaway is the trade the authors describe themselves: a cheap reasoning token in exchange for weaker recall of facts, which shifts work to retrieval rather than to a larger model.
Source: Hugging Face blog (Liquid AI), "Run fast, private agents on-device with LFM2.5-24B-A2B" — https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b