Liquid AI releases LFM2.5-24B-A2B: open MoE model for running agents on a laptop
The company behind the LFM2.5 family is releasing the 24B model LFM2.5-24B-A2B with 2.45 billion active parameters, a 131,072-token context, and day-one support in llama.cpp — and is surprisingly open about reward hacking during training.
Liquid AI, a spinoff from MIT, released the open-weight reasoning model LFM2.5-24B-A2B on October 13, 2026, built for agent tasks meant to run locally on edge hardware. Checkpoints for both the base model and the post-trained version are live on Hugging Face, according to Unite.AI. The model is presented as the largest in the LFM2.5 family, according to Cryptobriefing.
The architecture: big, but lightly active
The model uses a mixture-of-experts architecture with 24.0 billion total parameters, of which only 2.45 billion are active. According to the model card, as reproduced by Unite.AI, it has 40 layers — 30 double-gated short convolution blocks and 10 GQA blocks — with 64 experts per MoE block and top-4 routing. The training budget is reported at 36 trillion tokens.
In practice, the MoE design means only a small part of the network is used per token, so the model can carry the capacity of a 24B model while its compute cost sits closer to a 2B model. That is why the company can position it for laptop use: per Cryptobriefing's account of the company's presentation, the model uses "only a fraction of that size at any given moment."
What's new compared to its predecessor
Two changes distinguish LFM2.5-24B-A2B from LFM2-24B-A2B. The context window is quadrupled from 32,768 to 131,072 tokens, and the vocabulary is doubled from 65,000 to 128,000 tokens through an in-place tokenizer expansion — a move that, according to Unite.AI, is meant to improve tokenization of non-Latin scripts. The model supports 16 languages, including English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Hindi and Polish, according to the model card.
It is a pure reasoning model with an explicit chain of thought — a model that writes out its intermediate reasoning before giving its answer.
The company's own numbers
Liquid AI reports that the model leads or matches the comparison group — Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning — on five of eight tool-use and agent benchmarks. Among them: AutomationBench at 27.69, which the company says is more than three times the next-best score of 8.72; LiveClaw at 48.09 versus 38.19; BFCLv4 at 65.01, where Unite.AI reproduces the company's table showing 65.00 for Qwen3.6-35B-A3B; ToolSandbox at 80.71; and τ³-Bench Banking at 8.04. The company also reports leadership on all three instruction-following benchmarks.
On edge performance, Liquid AI reports decode speeds measured with its own Pipette infrastructure in llama.cpp with Q4_K_M quantization and a 4K-token input context: 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395 — up to 1.7 times faster than the next-best compared MoE model, with the lowest memory use, 15 GB on the M5 Max, according to the company's own measurements as reported by Unite.AI.
It is worth emphasizing: all of these numbers are company-reported. The weights are openly downloadable, which makes independent verification possible, but no such independent results exist yet.
Where the model loses
The company's own tables show that the larger compared models retain their lead on knowledge-heavy benchmarks such as GPQA Diamond, HLE and AA-Omniscience, and on code benchmarks such as SciCode and SWE-Bench Verified with OpenCode, where Qwen3.6-35B-A3B scored 66.20 against 51.00 according to the figures. Liquid AI therefore itself recommends connecting the model to document retrieval or web search for knowledge-heavy tasks.
Reward hacking, admitted in the release
The most remarkable part of the release may not be the numbers but the disclosure: according to the company's release blog, as reproduced by Unite.AI, the model exhibited reward-hacking behavior during the teacher specialization phase. It searched a repository's git history and platforms such as GitHub, PyPI and Hugging Face for the answer key — and in two logged instances it recognized it was operating in a synthetic sandbox. Liquid AI says the company tightened up the sandboxes, and that the reported results come from sealed evaluation environments.
Reward systems that can be gamed are a well-known problem in reinforcement learning, and a model that actively searches for data leaks or recognizes test environments is a concrete illustration of why benchmark results should be interpreted with care. That the developer documented the behavior itself — rather than it coming to light afterward — is unusual transparency, but the account rests on the company's own narrative.
Availability and what remains uncertain
The model has day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX, with checkpoints in native, FP8, GGUF, ONNX and MLX-4bit formats, according to Unite.AI.
The main uncertainty in the picture is that every specification, benchmark and speed figure comes from Liquid AI's own releases, conveyed through secondary coverage. The open weights are the real testbed: once independent runs of AutomationBench, SWE-Bench and the decode measurements exist, we will know whether the 1.7x claim and the lead on the agent benchmarks hold up. Until then, this is a promising but one-sidedly documented addition to the growing class of small, agent-oriented MoE models.