Liquid AI announces LFM2.5-24B-A2B, a 24-billion-parameter reasoning model where only 2 billion parameters are active per token, designed to run fast, private agents directly on a device. The model thinks before answering, calls tools, and chains multi-step tasks within the most popular agent environments. Its open weights are available on Hugging Face.
The model retains the architecture of its predecessor, LFM2-24B-A2B, with 40 layers, 64 experts per MoE block (mixture of experts, an architecture type where only a portion of the network is activated per computation), and top-4 routing. The context window increases from 32K to 131K tokens, and the vocabulary expands from 65K to 128K tokens to better tokenize non-Latin scripts. Unlike LFM2-24B-A2B, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer.
Pre-training was scaled from approximately 17T to about 33T tokens. This was followed by intermediate training, with 2T tokens at a 32K context, focusing on reasoning, mathematics, code, instruction following, and multilingual texts, then 300B tokens at a 131K context. Post-training transforms this base model into an agent in five steps: supervised fine-tuning (SFT), specialization of ‘teacher’ models, two phases of Mixture of Policy Distillation (MOPD)—one on non-agentic tasks, the other on agentic tasks—and finally agentic reinforcement learning, conducted in agent environments like OpenClaw, Hermes Agent, Pi, and OpenCode.
Reported Results Against Larger Models
Liquid AI compared its model to Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning on 16 benchmarks covering tool use, knowledge, instruction following, mathematics, and code. According to the company, LFM2.5-24B-A2B is the smallest in the group, yet remains competitive against models with roughly 50% more total and active parameters. On LiveClaw, it scores 48.09 versus 38.19 for Qwen3.6-35B-A3B. On AutomationBench, it reaches 27.69, compared to 8.72 for the best competitor in the published chart. It also leads the other models on the three instruction-following benchmarks, including IFEval (96.75) and IFBench (74.69). Larger models retain the advantage on tests requiring broad knowledge, such as GPQA Diamond, where Qwen3.6-35B-A3B scores 83.69 versus 74.42, as well as on code, for example SWE-Bench Verified (OpenCode), where it gets 66.20 versus 51.00. Liquid AI therefore recommends pairing this model with document retrieval or web search for knowledge-heavy use cases. Evaluations were conducted with vLLM, using the recommended generation parameters for each benchmark.
Regarding hardware performance, the company reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, which is up to 1.7 times faster than the next model, while consuming the least memory. On a single NVIDIA B300 GPU, the model achieves up to 21K output tokens per second under high concurrency.
The model is compatible out of the box with llama.cpp, MLX, vLLM, SGLang, and ONNX. To use it with an agent, simply serve it locally behind an OpenAI-compatible interface, then point the agent to this server’s address, for example on port 8080 for llama.cpp and MLX, or port 8000 for vLLM.
