Each model is drawn against the November 2026 LLM blueprint: which components it keeps and how it configures them, what it does differently, and what it leaves out. Every profile cites the model's own report, config or code.
Compare two models side by sideDeepSeek, September 2026
552B + 196B Engram total, 16B (8B in prefill) active, 1M tokens
Causal encoder-decoder MoE: decoder reads encoder-projected KV, CSA2 shares KV across layers, FP4 KV at 890 B/token, Engram, DSpark.
Of 27 blueprint components: 23 same, 3 different, 1 not used.
Thinking Machines Lab, July 2026
975B total, 41B active, 1M tokens (1,048,576)
5:1 sliding-window/global GQA with a learned relative-position bias instead of RoPE, short causal convs, encoder-free image and audio.
Of 27 blueprint components: 18 same, 5 different, 3 not used, 1 not disclosed.
Moonshot AI, July 2026
2.8T total, 104B active, 1M tokens (1,048,576)
93 layers of 3:1 KDA / gated MLA with no positional encoding, Block Attention Residuals and a 16-of-896 latent MoE in MXFP4.
Of 27 blueprint components: 19 same, 5 different, 3 not used.
NVIDIA, June 2026
550B total, 55B active, 1M tokens
Mamba-2 does most of the token mixing, 12 NoPE GQA layers anchor it, and LatentMoE routes 22 of 512 experts in a 2,048-dim latent.
Of 27 blueprint components: 16 same, 4 different, 7 not used.
DeepSeek, April 2026
1.6T total, 49B active, 1M tokens
Attention over a compressed sequence: CSA (4→1, top-1,024) and HCA (128→1) layers with a 128-token window, 4-stream mHC, Muon.
Of 27 blueprint components: 22 same, 1 different, 3 not used, 1 not disclosed.
Google DeepMind, April 2026
30.7B, 256K tokens
Dense 5:1 local/global stack: 1,024-token windows, global layers with K = V and p-RoPE, sandwich norms, tied 262K vocabulary.
Of 27 blueprint components: 10 same, 7 different, 6 not used, 4 not disclosed. Adds 2.
MiniMax, March 2026
229.9B total, 9.8B active, 204,800 tokens (M2 to M2.5: 196,608)
Full GQA attention in all 62 layers after hybrid-attention ablations; 256-expert top-8 sigmoid MoE with 9.8B active; 3 MTP modules.
Of 27 blueprint components: 14 same, 4 different, 7 not used, 2 not disclosed.
Z.ai (Zhipu AI), February 2026
744B total, 40B active, 202,752 tokens (GLM-5.2: 1,048,576)
MLA with DeepSeek Sparse Attention in every layer, 256-expert MoE, one MTP layer reused for 3 steps; GLM-5.2 adds IndexShare and 1M context.
Of 27 blueprint components: 19 same, 3 different, 5 not used.
Alibaba Qwen, February 2026
397B total, 17B active, 262K tokens (1.01M with YaRN)
15 × (3 Gated DeltaNet + 1 gated GQA) layers over 512 small experts: only 15 of 60 layers keep a KV cache.
Of 27 blueprint components: 19 same, 3 different, 3 not used, 2 not disclosed.
OpenAI, August 2025
117B total, 5.1B active, 131K tokens
A deliberately plain MoE: alternating 128-token banded and dense GQA layers, attention sinks, MXFP4 experts.
Of 27 blueprint components: 12 same, 4 different, 9 not used, 2 not disclosed.
A dot means the model uses the blueprint's design (hover for its configuration). Named cells are where it does something else. A dash means it doesn't use the component at all; a question mark means its authors haven't said.