SGLang provides day-0 support for the model, which reaches output speeds of 21K tokens per second on a single NVIDIA B300 and 184 tokens per second on an Apple M5 Max. Liquid AI developed the model for multi-step agenting tasks within popular harnesses, prioritizing local execution to reduce computational costs and improve privacy.
Tuesday, October 20, 2026
Story
Liquid AI Launches 24B Model as Largest Release to Date
AI7 days ago
The new reasoning tool utilizes a Mixture-of-Experts (MoE) architecture to enable fast, private agentic intelligence on laptops and PCs. The LFM2.5-24B-A2B model consists of 24B total parameters with 2B active per token, a 131K token context window, and a 128K token vocabulary. Trained on 33T tokens, the model has open base and post-trained weights on Hugging Face and outperforms the Qwen3.6-35B-A3B on agentic tasks, scoring 48.09 vs 38.19 on LiveClaw and 27.69 vs 8.72 on AutomationBench.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.