This is a direct sequel to the article “meta-attention is all you need”. Reading it first is recommended but not required — a tour of the architecture is included below.

The first article described the idea in detail and published the experiment sources, but “sources” there meant simply all the code that had been written, with no system to it — including several different implementations of the same components.

So here is a framework with a ready-made toolkit you can try out on LLMs, including in agentic scenarios.

Ready-to-use, lightweight trained wrappers are also provided — one tiny model (Qwen-3.5–4B) and one mid-size one (Granite 3.3 8B). All of them can be run through llama.cpp.
https://medium.com/@artem-x/building-the-meta-spider-framework-on-top-of-meta-attention-476eb95bb85b?postPublishedType=initial