Physix FrontierPhysix Frontier · News Briefing Card (HuggingFace · Oct 14, 2026)
Liquid AI Releases LFM2.5-24B-A2B On-Device Reasoning Agent Model
KEY FACTS
- Liquid AI released LFM2.5-24B-A2B, with 24B total parameters and 2B activated per token.
- The model is reasoning-specific: it first outputs a chain of thought, then gives the final answer.
- The context window expands from 32K to 131K, and the vocabulary expands from 65K to 128K.
- The pretraining token count rises from about 17T to about 33T.
- It scores 48.09 on the LiveClaw benchmark, versus 38.19 for Qwen3.6-35B-A3B.
KEY DATA
24BTotal Parameters
2BActivated Parameters
48.09LiveClaw Score
184 tok/sApple M5 Max Inference Speed
PHYSIX OBSERVATION
On-device agents are moving from can run to can run continuously. With 2B activated parameters, the cost per inference is pushed extremely low, while 24B total parameters preserve tool-calling capability. This combination is more noteworthy than simply piling on parameters. For users, it means privacy and response speed can be had at the same time; for the industry, it shows that MoE sparsification plus reasoning chains is becoming the default route for on-device agents.
Source: HuggingFace report