Liquid AI releases LFM2.5-24B-A2B for on-device agents

Liquid AI released LFM2.5-24B-A2B on October 13, 2026.[1] It describes the model as having 24 billion parameters, with 2 billion active.[2] Liquid AI says it reasons before producing its final answer, while the Hugging Face post describes it as able to think before answering, call tools and work through multi-step tasks inside popular agent harnesses.[2][3] The company calls it the largest and final model in the LFM2.5 family, whose models range from 230 million to 24 billion parameters.[2] It is open-weight and available on Hugging Face and Liquid AI’s Playground.[2]

Liquid AI increased the context window from 32,768 to 131,072 tokens and doubled the vocabulary from 65,000 to 128,000 tokens, which it says is intended to improve non-Latin tokenization.[2] The company also scaled pretraining from 17 trillion to 33 trillion tokens.[2] Its first mid-training phase used 2 trillion tokens at a 32,000-token context, emphasizing reasoning, math, code, instruction following and multilingual text.[2] A second phase used 300 billion tokens at a 131,000-token context and weighted the data toward long-context material.[2]

Liquid AI says it evaluated the model on 16 benchmarks covering agent workflows, knowledge, instruction following, math and coding.[2] It reports that the model led or matched competitors on five of eight tool-use and agentic benchmarks, including scores of 48.09 on LiveClaw versus 38.19, and 27.69 on AutomationBench versus 8.72.[2] The company says the model led all three instruction-following benchmarks in its comparison.[2] It also reports that larger models did better on knowledge-heavy benchmarks such as GPQA Diamond, HLE and AA-Omniscience, and recommends retrieval or web search for those tasks.[2]

Liquid AI reports decode speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395.[2] It says the model is up to 1.7 times faster than the next-best model while using the least memory, and reports throughput above 21,000 output tokens per second at high concurrency.[2] A company demonstration ran the model entirely on-device without cloud API calls.[2] For hosted inference, Liquid AI lists prices of $0.10 per million input tokens, $0.05 per million cached input tokens and $1 per million output tokens, plus $2 in starter credit.[2]