Every modern LLM architecture — 119 models from DeepSeek-V3 to Llama 4, Qwen3, Kimi K2 and gpt-oss — rendered as one interactive 3D atlas. Every layer, every head, every routing decision. Mapped.
$ atlas inspect deepseek-v3 ▸ 671B MoE · 256 experts · MLA attention $ atlas compare qwen3-32b llama-4-maverick ▸ dense vs MoE · GQA vs chunked attention · rendering… ◇ 119 architectures loaded — fact-checked, source-linked
Attention blocks, MoE routers, norm placement — walk each architecture layer by layer in 3D instead of squinting at block diagrams.
MLA vs GQA, dense vs sparse, RoPE variants and vocabulary trade-offs — side-by-side, from the same verified fact sheets.
Every field passes a schema, source and dead-link gate. Nothing enters the atlas unverified. No hype — just signal.