The Open Weights Cascade — 25+ AI Model Drops in One Week Across Every Modality

This week may be the most significant seven days in open-source AI history. As Victor M (@victormustar) documented in a viral thread 25+ notable open-weight models dropped across every modality — LLMs, multimodal, audio, video, image, code, and reasoning.

The cascade is not random. It’s structural.

What Dropped This Week

Nvidia Nemotron 3 Ultra — 550B parameters, hybrid Mamba-MoE with only 55B active, 1M context, MMLU 89.1. Nvidia giving away a frontier-class reasoning model. The company that sells GPUs is now giving away the models that run on them.

Google Gemma 4 12B — first medium-sized model with native audio, video, image, and text. Encoder-free. Runs on 16GB VRAM. Apache 2.0. Google doing to multimodal AI what Android did to mobile.

Plus releases from Meta (Llama), Mistral, Qwen, and dozens of specialized models across code, speech, image, and video.

Why It’s a Cascade, Not a Coincidence

GPU vendors subsidize open models. Nvidia releases Nemotron because every developer who adopts it buys Nvidia GPUs. The model is the razor. The GPU is the blade.

Cloud providers commoditize the model layer. Google releases Gemma because free models increase demand for Google Cloud compute. Meta releases Llama because open models run on PyTorch on Meta’s tooling.

The frontier is compressing. Open models at 90% of frontier quality capture 90% of use cases — at zero licensing cost. The proprietary advantage erodes every week.

What This Means

Winners in a commoditized model world: infrastructure providers (Nvidia, TSMC), orchestration platforms (Microsoft, Salesforce), and vertical AI companies (domain-specific applications where the model is an input, not the product).

Losers: pure-play model providers who can’t differentiate on capability alone. 25+ models in one week. The cascade is accelerating.

Sources

• Victor M (@victormustar) — open weights cascade thread
• Google Gemma 4 — Google Blog

Get structural AI business analysis every week — free. Subscribe to Business Engineer →