Liquid AI releases a downloadable AI model meant to run on your own device
Liquid AI says its new model can work through multi-step tasks on your own hardware instead of in the cloud.
Liquid AI says the model thinks before it answers and can use outside tools to finish tasks with several steps. It is designed to run on your own computer, not on a cloud service.
The model has 24 billion adjustable settings, but Liquid AI says the model uses only about 2 billion at a time, which keeps each step cheap and fast. On an Apple M5 Max chip, Liquid AI says the model writes 184 tokens (small chunks of words) per second. That is up to 1.7 times faster than the next-best model it compared.
In Liquid AI's own tests, the model scored 48.09 on a tool-use test called LiveClaw. Qwen3.6-35B-A3B, a model about 50% larger, scored 38.19. Liquid AI says larger models still win on knowledge and coding tests.
Official releaseRead the original at Hugging Face →Developers can download the model today and run it on their own hardware. Liquid AI says larger models still do better on knowledge-heavy and coding tests.
The facts
- LFM2.5-24B-A2B has 24 billion parameters in total, but only 2 billion are active for each token it processes.
- Liquid AI says the model scored 48.09 on LiveClaw, versus 38.19 for Qwen3.6-35B-A3B, a model with about 50% more total and active parameters.
- Liquid AI says the model decodes 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395.
- Liquid AI says that is up to 1.7 times faster than the next-best model it compared, while using the least memory.
- Liquid AI compared the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on 16 tests.
- Liquid AI says larger models keep an edge on knowledge-heavy tests (GPQA Diamond, HLE, AA-Omniscience) and coding tests (SWE-Bench Verified, SciCode).
- The model and its base version are open-weight and available to download on Hugging Face today.
Sources
- Hugging Face: Run fast, private agents on-device with LFM2.5-24B-A2B
Published
More in New models
Google releases Gemini 4 Argon, its newest AI, to paying users first
Google says its new AI matches the best from rivals. Pro and Ultra subscribers get it today, while free users have no date.
New modelsBig storyMeta's new AI models reach the top tier, but cost nearly three times more than before
Meta's Muse Wonder models score close to the best from OpenAI and Anthropic in tests. Running them costs nearly three times what Meta's older model did.