SGLang announces Day-0 support for LFM2.5-24B-A2B, a 24B-parameter mixture-of-experts reasoning model from Liquid AI with 2B active parameters and a 131K context window.
The post reports 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench against Qwen3.6-35B-A3B, and up to 21K output tokens per second on a single NVIDIA B300.
Liquid AI's post says the open base and post-trained weights are available on Hugging Face.
SGLang has Day-0 support for LFM2.5-24B-A2B from @liquidai
A 24B MoE reasoning model with a 131K context window. Built for fast, private agents on-device.
✅ Competitive with Qwen3.6-35B-A3B on agentic and tool-use tasks: 48.09 vs 38.19 on LiveClaw, 27.69 vs 8.72 on AutomationBench
✅ Up to 21K output tok/s on a single NVIDIA B300
Get started with SGLang:
Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest 𝙻𝙵𝙼𝟸.𝟻 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses. > 24B total parameters, 2B active per token > Context expanded to 131K tokens > A 128K-token vocabulary > 33T pretraining tokens > It is a reasoning model > Very easy to fine-tune and customize > Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks. > Open base and post-trained weights on @huggingface 🤗 1/
Source: SGLang · x.comPublished
