Liquid AI Releases LFM2.5-24B-A2B for Fast, Private Local Agents
Liquid AI’s new 24B-parameter mixture-of-experts model activates 2.45B parameters per token, supports a 131K context window and ships in multiple local-deployment formats.
2 min read
Illustrative photo by Stefan Cosma on Unsplash; not a Liquid AI product image.
Liquid AI released LFM2.5-24B-A2B on October 13, adding a downloadable mixture-of-experts reasoning model aimed at private, local agents. The model has 24 billion total parameters but activates 2.45 billion per token, with 64 experts and top-four routing. It is text-only, supports a 131,072-token context window and lists 16 supported languages.
The canonical Hugging Face repository provides the post-trained checkpoint for Transformers, vLLM and SGLang, alongside links to FP8, GGUF, ONNX and MLX variants. Liquid says the Q4_K_M GGUF is 14.5 GB, making the model practical for some high-memory laptops and workstations. The weights use Liquid’s custom LFM Open License 1.0, so “open-weight” is more precise than claiming an OSI-approved open-source license.
Liquid reports that the model scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and claims throughput of 184 tokens per second on an Apple M5 Max. Both are vendor-reported measurements; performance will vary with quantization, runtime, hardware and agent harness. The model card also cautions that the checkpoint is not designed to be a standalone knowledge-heavy assistant and may require retrieval or tools.
The release is notable because it combines a larger total parameter budget with a small active footprint and day-one local-serving support. It gives developers a new option for on-device tool use, but independent evaluations are still needed to test the vendor’s speed and agent claims across reproducible setups.
Sources
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.