24 billion parameters, 2 billion active
Liquid AI released LFM2.5-24B-A2B on Hugging Face today. It is a reasoning model with 24 billion parameters and 2 billion active per token. The architecture follows the previous version: 40 layers, 64 experts per mixture-of-experts block, and top-4 routing. Context grew from 32,000 tokens to 131,000, and the vocabulary from 65,000 to 128,000. Pre-training grew from about 17 trillion tokens to about 33 trillion.
Both the instruction model and the base weights are public. It writes a chain of thought before the answer.
[1]
Tool use matches larger models. Coding does not
The comparison set is Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning. Liquid AI says this is the smallest model in the group, and that it matches or leads on five of eight tool-use and agent tests: AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. LiveClaw is 48.09 against Qwen's 38.19. AutomationBench is 27.69, and the next score is 8.72.
On the same table, SWE-bench Verified with OpenCode is 51.00 against Qwen's 66.20. GPQA Diamond is 74.42 against 83.69. They say knowledge and coding still lag, and they suggest retrieval or web search for knowledge-heavy work.
[1]The speeds are measured on their machines
The post reports 184 tokens per second on an Apple M5 Max, and more than 21,000 output tokens per second at high concurrency on one NVIDIA B300. They call it the fastest of the compared mixture-of-experts models, with the least memory. That is the publisher's measurement, not a third-party rerun.
[1]