LLM Architecture Gallery
https://sebastianraschka.com/llm-architecture-gallery/On the LLM Architecture Gallery, it’s interesting to see the variations between models, but I think the 30,000ft view of this is that in the last seven years since GPT-2 there have been a lot of improvements to LLM architecture but no fundamental innovations in that area. The best open weight models today still look a lot like GPT-2 if you zoom out: it’s a bunch of attention layers and feed forward layers stacked up.
Another way of putting this is that astonishing improvements in capabilities of LLMs that we’ve seen over the last 7 years have come mostly from scaling up and, critically, from new training methods like RLVR, which is responsible for coding agents going from barely working to amazing in the last year.
That’s not to say that architectures aren’t interesting or important or that the improvements aren’t useful, but it is a little bit of a surprise, even though it shouldn’t be at this point because it’s probably just a version of the Bitter Lesson.
After years of showing up in papers and toy models, hybrid architectures like Qwen3.5 contain one such fundamental innovation - linear attention variants which replace the core of transformer, the self-attention mechanism. In Qwen3.5 in particular only one of every four layers is a self-attention layer.
MoEs are another fundamental innovation - also from a Google paper.
Is there a sort order? Would be so nice to understand the threads of evolutions and revolution in the progression. A bit of a family tree and influence layout? It would also be nice to have a scaled view so you can sense the difference in sizes over time.
I even brought my popcorn :(
And there are still plenty of hybrid architectures. Nemotron 3 Super 120B A12B just came out, it's mostly Mamba with a few attention layers, and it's pretty competitive for its size class.
But yeah, these different architectures seem to be relatively small micro-optimizations for how it performs on different hardware or difference in tradeoffs for how it scales with the context window, but most of the actual differentiation seems to be in training pipeline.
We are seeing substantial increases in performance without continuing to scale up further, we've hit 1T parameters in open models but are still having smaller models outperform that with better and better training pipelines.
Right now we’re engineering every bit of it to make it better but in the long run this is unsustainable. It’s going to be so complex that even these digital life forms won’t be able to understand their own digital DNAs, like us.
We know we have DNA, we can measure every letter but it doesn’t mean we understand what’s going on our 14 trillion cells and how each and every one of them is regulated.
I think this analogy not only explains us, or digital beings we see today. It explains everything, quite literally. Still it would be amazing to think about these systems from the perspective of biology, and try to understand the parts analogous to existing frame that we already have. Then we might figure out what to optimize better. For instance if we figure out a certain part of a layer corresponds to “genes” then we might find out there is alternative splicing within it. Wild but worth a shot.