Evidence-traceable LLM architecture atlas

Catalog

6 models. Every figure is generated from an Architecture IR; every number traces to an evidence claim.

DeepSeek V3 (671B) DeepSeek · MLA + MoE 671B total / 37B active · 61 layers · 160K ctx
GLM-5.3-Flash (320B-A18B) GLM · Hybrid (KDA + MLA/DSA) + MoE 320B total / 18B active · 45 layers · 1M ctx
Kimi Linear 48B-A3B Kimi · Hybrid (KDA + MLA) + MoE 48B total / 3B active · 27 layers · 1M ctx
Llama 3 8B Llama · GQA 8B total · 32 layers · 8K ctx
Mixtral 8x7B Mixtral · GQA + MoE 46.7B total / 12.9B active · 32 layers · 32K ctx
Qwen3-Next 80B-A3B Qwen3-Next · Hybrid (Gated DeltaNet + Gated Attention) + MoE 80B total / 3B active · 48 layers · 256K ctx