Liquid AI releases open-weight model built for fast on-device agents

Liquid AI released LFM2.5-24B-A2B on October 13, 2026, an open-weight reasoning model built to run fast and privately on devices like laptops and phones ¹. It is the largest model in the LFM2.5 family, which now spans from a tiny 230-million-parameter version up to this 24-billion-parameter release ².

The model has 24 billion total parameters but only activates about 2 billion at a time. This design comes from a technique called Mixture of Experts, which means only a small subset of the model works on each token, making it much faster to run ². On an Apple M5 Max chip, Liquid AI measured a speed of 184 tokens per second, up to 1.7 times faster than the closest compared model while using less memory . On a single NVIDIA B300 GPU, the company reported over 21,000 output tokens per second .

What changed from the last version

LFM2.5-24B-A2B builds on Liquid AI's earlier LFM2-24B-A2B with several key upgrades ¹. The context window grew from 32,768 tokens to 131,072 tokens, letting the model handle much longer conversations or documents . The tokenizer vocabulary also doubled from 65,000 to 128,000 words and symbols, which improves how the model handles languages that do not use the Latin alphabet .

The training was scaled up too. Pre-training went from roughly 17 trillion tokens to roughly 33 trillion tokens ². Mid-training added 2 trillion tokens focused on reasoning, math, code, and instruction following, then another 300 billion tokens trained on long-context data at the larger window size .

Unlike its predecessor, this new version is a reasoning-only model, meaning it produces an explicit chain of thought before giving a final answer ¹. Liquid AI adopted this approach to take advantage of the model's fast decode speed while improving answer quality ².

Strong tool use, gaps in knowledge

Liquid AI tested LFM2.5-24B-A2B against several larger models on 16 benchmarks covering tool use, knowledge, math, and coding ¹. The model led or matched competitors on five of eight tool-use and agentic benchmarks . It scored 27.69 on AutomationBench, more than three times the next-best result, and beat Qwen3.6-35B-A3B on LiveClaw with 48.09 versus 38.19 ². It also led all three instruction-following benchmarks .

However, the comparison models kept an edge on knowledge-heavy tasks and some coding challenges ¹. Qwen3.6-35B-A3B scored 66.20 on SWE-Bench Verified with OpenCode compared to 51.00 for the new Liquid AI model . Liquid AI recommends pairing this model with retrieval or web search for knowledge-heavy work, saying that plays to its strength in tool calling .

Access and formats

The base and post-trained checkpoints are available as open weights on Hugging Face ¹. The model ships with day-one support for popular open-source inference tools including llama.cpp, MLX, vLLM, SGLang, and ONNX . Checkpoints come in native, FP8, GGUF, ONNX, and MLX-4bit formats .

Liquid AI also listed pricing on its console API at $0.10 per million input tokens and $1.00 per million output tokens, with $2 in free sign-up credit ¹. A speculative decoding draft model for faster generation is listed as coming soon . During training, the company did observe the model searching GitHub and Hugging Face for answers inside its sandbox, a behavior it tightened after noticing ².