This is great - always worth reading anything from Sebastian. I would also highly recommend his Build an LLM From Scratch book. I feel like I didn’t really understand the transformer mechanism until I worked through that book.
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.
> 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.
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.
Thanks for the note about Qwen3.5. I should keep up with this more. If only it were more relevant to my day to day work with LLMs!
I did consider MoEs but decided (pretty arbitrarily) that I wasn’t going to count them as a truly fundamental change. But I agree, they’re pretty important. There’s also RoPE too, perhaps slightly less of a big deal but still a big difference from the earlier models. And of course lots of brilliant inference tricks like speculative decoding that have helped make big models more usable.
I'd push back slightly on the "no fundamental innovations" read though — the innovations that stuck (MoE, GQA, RoPE) are almost entirely ones that improve GPU utilization: better KV-cache efficiency, more parallelism in attention, cheaper to serve per parameter. Mamba and SSM-based hybrids are interesting but kept running into hardwar friction.
This is amazing, such a nice presentation. It reminds me of the Neural Network Zoo [1], which was also a nice visualization of different architectures.
Thank you for this! I help teach a "CS enrichment course", and I'm having students play with Keras (with my own written scaffolding of course.) I'm struggling to find a resource to help me plan beyond "this is a perceptron/FFNN", and with my lack of experience (I'm a statistician) this is going to be extremely helpful.
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.
> The goal of the proof verifier (LLM 2) is to check the generated proofs (LLM 1), but who checks the proof verifier? To make the proof verifier more robust and prevent it from hallucinating issues, they developed a third LLM, a meta-verifier.
The one thing I didn't quite understand (and wasn't mentioned in their paper unless I missed it), is why you can't keep stacking turtles. You probably get diminishing returns at some point, but why not have a meta-meta-verifier?
This is amazing, I just spent some time scrolling through these, most of the evolution is about inference cost not capability. GQA, MoE routing, sliding window attention, all trading theoretical capacity for practical efficiency.
Tbh might be the last generation of architectures designed entirely by humans. I dug into that (1) and might add another paragraph based on this if I find the time. The Big LLM Architecture Comparison (2) by Sebastian Raschka already inspired my ograph image for the blog -thanks again!
Nice! Last time I had a custom temporary tattoo made, I had to copy and paste from Attention is All You Need; this provides a much cleaner and more varied source.
It is perhaps my eyes, but when I zoom in enough to make it readable, it gets blurry. A higher-res image would be much appreciated. Great idea otherwise.
Thank you so much! As a (bio)statistician, I've always wanted a "modular" way to go from "neural networks approximate functions" to a high-level understanding about how machine learning practitioners have engineered real-life models.
Interesting collection. The architecture differences show up in surprising ways when you actually look at prompt patterns across models. Longer context windows don't just let you write more, they change what kind of input structure works best.
Competitiveness doesn't really come from architecture, but from scale, data, and fine-tuning data. There has been little innovation in architecture over the last few years, and most innovations are for the purpose of making it more efficient to run training or inference (fit in more data), not "fundamentally smarter"
If your definition of "competitive" is loose enough, you can write your own Markov chain in an evening. Transformer models rely on a lot of prior art that has to be learned incrementally.
Where are you seeing dense? Most of the larger competitive models are sparse. Sure, the smaller models are dense, but over 30B it's pretty much all sparse MoE.
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.
41 comments
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.
> 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.
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.
I did consider MoEs but decided (pretty arbitrarily) that I wasn’t going to count them as a truly fundamental change. But I agree, they’re pretty important. There’s also RoPE too, perhaps slightly less of a big deal but still a big difference from the earlier models. And of course lots of brilliant inference tricks like speculative decoding that have helped make big models more usable.
[1] https://www.asimovinstitute.org/neural-network-zoo/
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.
> The goal of the proof verifier (LLM 2) is to check the generated proofs (LLM 1), but who checks the proof verifier? To make the proof verifier more robust and prevent it from hallucinating issues, they developed a third LLM, a meta-verifier.
Tbh might be the last generation of architectures designed entirely by humans. I dug into that (1) and might add another paragraph based on this if I find the time. The Big LLM Architecture Comparison (2) by Sebastian Raschka already inspired my ograph image for the blog -thanks again!
(1) https://philippdubach.com/posts/the-last-architecture-design...
(2) https://magazine.sebastianraschka.com/p/the-big-llm-architec...
I even brought my popcorn :(
I’m thinking it’s still llama / dense decoder only transformer.
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.