As long as the weights aren't identical, that's fine by me... all inference optimizations should still work then, which is actually a benefit 🚀

People really need to stop living in the fantasy world of corporate marketing guidelines.

The "scary truth" is that almost all modern LLMs crawled out of the exact same place. The foundational architecture has been practically frozen across the industry for a long time—whether you are looking at a cloud giant like Claude or local models like Qwen.

When teams release a "new" incremental version, nobody is wasting millions to train a massive base from scratch. They take the frozen core, tweak the attention heads (which mostly just alters processing speed and token throughput), and do fine-tuning on a few specific layers.

It’s the same old engine with a fresh coat of paint and some minor tuning under the hood. Of course the architecture diff shows 0 changes—it's just a glorified fine-tune of the same base.