Overview
- Liquid AI published LFM2.5-24B-A2B on Hugging Face on Tuesday, October 13, 2026, and both the full model and a base variant are open-weight and available for download.
- The model uses a mixture-of-experts design that holds 24 billion parameters but activates about 2 billion parameters per token to reduce runtime cost, and it produces an explicit chain of thought before answering.
- LFM2.5-24B-A2B expands context and token handling by raising the context window to 131,000 tokens and doubling the tokenizer vocabulary to 128,000 entries to help with long documents and non-Latin text.
- Liquid AI published vendor-run benchmarks and speed figures showing strong agentic and tool-use scores versus larger rivals and reporting edge inference speeds (for example, 184 tokens/s on an Apple M5 Max and over 21K output tokens/s on an NVIDIA B300), but those results were measured in-house and not independently verified.
- The release includes day-one support for popular backends and agent harnesses so teams can serve the model locally, and Liquid AI recommends pairing it with retrieval or web search for knowledge- and code-heavy tasks because larger models still lead on some benchmarks; pricing and license terms were not disclosed.