The open model ecosystem has fragmented into at least three distinct maker categories, and none of them look like the Llama era. A year ago, Chinese labs dominated open releases. Now the field includes sovereign AI players like Mistral and Trillion Labs, product companies like JetBrains and Photoroom training small specialized models to avoid dependence on closed APIs, and Big Tech players with explicit hardware or upsell motives. Alibaba releases Qwen to drive closed-model revenue. NVIDIA releases Nemotron to sell GPUs. The motivations are legible now in a way they were not before.
The headline model this round is NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, a 550-billion-parameter sparse model using LatentMoE for speed advantages over dense comparables. Majority of training data is open source. More importantly, NVIDIA is dropping its custom license and adopting OpenMDW 1.1, a Linux Foundation license built specifically for model weights. MIT and Apache are software licenses. They do not cleanly cover weights. OpenMDW does. That licensing shift is worth understanding in detail, and the full piece explains exactly why it matters for downstream users and builders.
The original article's core argument is that frontier chasing will shrink as a share of open model activity, replaced by a long tail of specialized, product-adjacent releases. If that hypothesis is correct, the question of who controls the frontier matters less than who controls the tooling, licensing, and distribution layers beneath it. Attempts to restrict open model development, the piece argues, concentrate power among the few who can still access closed systems. Read the full piece for the breakdown of each new release and the licensing analysis.
[READ ORIGINAL →]