Liquid AI has released LFM2.5-24B-A2B, a reasoning model with 24 billion total parameters, of which only about 2 billion are active for each token. It was built to run fast, private agents directly on the user's device. It is the largest and final model of the LFM2.5 family, which now spans two orders of magnitude, from a 230-million-parameter model up to this one.
Activating only a few parameters at a time is what makes the model practical locally. According to the company's published data, the model is competitive with Qwen3.6-35B-A3B, which has about 50% more total and active parameters, on agentic and tool-use tasks. On LiveClaw, the benchmark cited by the company, it reportedly scores 48.09 against 38.19 for the competitor. These are figures reported by Liquid AI and not yet independently verified.
The model always reasons before answering, so each reasoning token is cheap. In inference tests the company describes, LFM2.5-24B-A2B would run up to 1.7 times faster than comparable MoE models on-device, at 184 tokens per second on an Apple M5 Max and up to 21K output tokens per second on a single NVIDIA B300 GPU. These numbers also come from a single source and should be read as the producer's claims.
Image: Hugging Face
Technically, the model keeps the LFM2-24B-A2B architecture with two changes: the context window grows from 32K to 131K tokens, and the vocabulary doubles from 65K to 128K tokens to handle non-Latin scripts better. Pre-training uses about 33 trillion tokens, against 17 trillion for its predecessor, followed by a 2-trillion-token mid-training phase on curated data and reinforcement learning inside popular agent harnesses.
Liquid AI also says where the model is not suited: it is not recommended for knowledge-heavy tasks on its own and should be paired with retrieval or web search, which plays to its tool-calling strength. The model is available in several formats on Hugging Face, including GGUF for llama.cpp, MLX for Mac, FP8 and ONNX. A draft model for speculative decoding with SGLang is announced as coming soon.
Impact on the Italian market
For developers building local agents, the practical point is open weights combined with lower inference cost. Those who prefer an API can use the price the company lists: $0.10 per million input tokens and $1.00 per million output tokens, with the model also available on OpenRouter.
No comments yet. Be the first!