Overview
- Liquid AI published LFM2.5-24B-A2B on Tuesday, October 13, 2026, and made both the post-trained agent and a base version available for open-weight download on Hugging Face.
- The model holds 24 billion total parameters but activates roughly 2–2.45 billion parameters per token using an MoE layout with 40 layers, 64 experts per block, a top-4 router, a 131,072-token context window, and a 128K-token vocabulary.
- Liquid AI reports vendor-run benchmarks showing the model leads or ties on several agentic and tool-use tests (for example, AutomationBench and LiveClaw) while noting larger models perform better on some knowledge and coding benchmarks.
- The company claims fast edge and GPU throughput—up to 184 tokens/sec on Apple M5 Max, 114 tok/s on Ryzen AI Max+ 395, and over 21,000 output tok/s at high concurrency on an NVIDIA B300—and ships examples for llama.cpp, vLLM, MLX, SGLang and ONNX.
- Reporters stress those results are vendor-provided and advise pairing the model with retrieval or web search for fact-heavy tasks, checking license terms for commercial use, and independently validating speed and accuracy by downloading the open weights.