LeCun highlights 25+ open-weight releases

The sudden influx of over 25 open-weight models in a single week signals that the AI industry is aggressively shifting from a centralized, proprietary-dominated landscape to a highly decentralized ecosystem where open-weight alternatives match or exceed the velocity of commercial offerings.

• –**Architectural Shift:** Developers are looking beyond standard dense Transformers, with major releases showcasing hybrid Mamba-MoE structures (Nemotron 3 Ultra), sparse MoEs (Step-3.7-Flash), and flow-matching Diffusion Transformers (Ideogram 4).
• –**Multimodal Ubiquity:** Native any-to-any capability (text, image, audio, video) is becoming a default expectation rather than a premium feature, as demonstrated by Google's Gemma 4 12B.
• –**Efficiency and On-Device Optimization:** Releases like Liquid AI's LFM2.5 and JetBrains Mellum2 indicate that the race is not just about parameter scale, but about optimizing active parameters for edge deployment and developer tasks.
• –**Mindshare Warfare:** By releasing highly capable models under open licenses, companies are building massive developer ecosystems to lock in mindshare before proprietary platforms can build equivalent custom integrations.