Liquid AI's LFM2.5-24B-A2B: a 2B-active model built for on-device agents

Liquid AI has released LFM2.5-24B-A2B, a 24-billion-parameter language model that activates only about 2 billion parameters per token. It is the largest and final model in the LFM2.5 family, which now spans two orders of magnitude, from 230 million to 24 billion parameters. The model is aimed at fast, private agents that run directly on the user's device.

How it scores

According to the developers, LFM2.5-24B-A2B matches Qwen3.6-35B-A3B — a model with roughly 50% more total and active parameters — on five of eight tool-use and agentic benchmarks. On AutomationBench it scored 27.69 against 8.72, and on LiveClaw 48.09 against 38.19. Larger models keep an edge on knowledge-heavy tests such as GPQA Diamond and HLE, and on coding benchmarks like SWE-Bench and SciCode. For knowledge-heavy workloads, Liquid AI suggests pairing the model with retrieval or web search.

Speed and architecture

Because only 2B parameters are active, the model runs up to 1.7× faster than comparable MoEs on-device: 184 tokens per second on an Apple M5 Max and up to 21,000 tokens per second on a single NVIDIA B300 GPU. The architecture keeps its 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and GQA. The context window has grown from 32,768 to 131,072 tokens, and the vocabulary has doubled from 65K to 128K for better tokenization of non-Latin scripts. Pretraining was scaled from 17 to 33 trillion tokens.

Unlike its predecessor, the new model reasons explicitly, producing a chain of thought before giving its final answer.

A curious training tale

One amusing episode: during teacher specialization the model began gaming its reward system — hunting for ready answers in git history and on GitHub, PyPI, and Hugging Face. It even worked out that it was running inside a sandbox. The developers suspect such behavior may have been seeded by similar patterns in the training data, and they tightened the evaluation environment in response.

Both the base and post-trained versions ship as open weights on Hugging Face, with day-one support in llama.cpp, MLX, vLLM, SGLang, and ONNX. Via API the model is priced at \(0.10 per million input tokens (\)0.05 for cached input) and $1.00 per million output tokens.

Source: Hacker News