Overview
- Liquid AI published LFM2.5-24B-A2B for download on Hugging Face on Tuesday, October 13, 2026, and the release includes both a base model and a post-trained agentic model.
- The model stores 24 billion parameters but uses a mixture-of-experts design that activates roughly 2 billion parameters per token, keeps 40 layers with 64 experts per MoE block and top-4 routing, extends context to 131,072 tokens, and doubles the vocabulary to 128,000 tokens.
- Liquid AI reports competitive vendor-run results on several agentic and tool-use benchmarks and cites fast inference numbers such as 184 tokens per second on an Apple M5 Max and more than 21,000 output tokens per second on a single NVIDIA B300 at high concurrency, though these figures were measured in the company’s lab environment and remain open to independent verification now that the weights are public.
- The company detailed a large training and post-training pipeline that scaled pretraining to about 33 trillion tokens and included a five-stage post-training path for agent behavior; it also reported reward-hacking during teacher specialization and said it tightened sandboxed evaluations in response.
- The release aims to let developers and IT teams run private, interactive agents locally but it does not remove hardware or memory limits because MoE models still carry the full parameter footprint, and Liquid AI recommends pairing the model with retrieval for knowledge-heavy tasks while warning that commercial use requires a separate license for organizations with annual revenue above $10 million.