LFM2.5-24B
A large language model released by Liquid AI that uses a MoE architecture with 24B total parameters, performs strongly on agentic and tool-use tasks, and is suitable for local deployment.
Timeline (last 90 days)
LFM2.5-24B ran successfully with a 128K context window.
In the LiveClaw evaluation, LFM2.5-24B scored 48.09, compared with 38.19 for the Qwen model.
LFM2.5-24B outperforms Qwen on most benchmarks and is only slightly weaker on Web Arena and Pinch Bench.
Running LFM2.5-24B uses about 28GB of memory, with the model fully loaded and running OpenClaw background tasks.
With a 128K context window, the uploader demonstrated the ability to handle a 50,000+ token session, with the terminal showing about 51,000 tokens used and the model still generating responses stably.
All Facts (5)
LFM2.5-24B ran successfully with a 128K context window.
50%UnverifiedIn the LiveClaw evaluation, LFM2.5-24B scored 48.09, compared with 38.19 for the Qwen model.
50%UnverifiedLFM2.5-24B outperforms Qwen on most benchmarks and is only slightly weaker on Web Arena and Pinch Bench.
50%UnverifiedRunning LFM2.5-24B uses about 28GB of memory, with the model fully loaded and running OpenClaw background tasks.
50%UnverifiedWith a 128K context window, the uploader demonstrated the ability to handle a 50,000+ token session, with the terminal showing about 51,000 tokens used and the model still generating responses stably.
50%