Liquid AI released LFM2.5-24B-A2B, an open-weight 24B Mixture-of-Experts reasoning model with 2B active parameters designed for on-device agent deployment (MarkTechPost). The model achieved 48.09 on LiveClaw and decoded at 184 tokens per second on an Apple M5 Max, targeting developers building local reasoning systems (Hugging Face Blog).
Google made Gemini 4 Argon generally available on the Gemini API, AI Studio, and the Gemini app, with support for up to 1M output tokens (MarkTechPost). The sparse MoE model leads the Vals Index at 68.90% (MarkTechPost). Separately, Google DeepMind and GovTech Singapore demonstrated lightweight on-device Gemma models for conversational scam detection with zero cloud data egress (Google DeepMind).