What changed

Once a proposal has enough votes from members, an AI reviewer checks the cited evidence, the technical claims and whether the change reflects the frontier. Approved proposals are merged into the live blueprint as a commit; each blueprint now has a history page, and any commit can be reverted, like git revert. Proposals the reviewer doesn't approve stay open for discussion and votes, with the review attached.

Each model page shows the LLM blueprint with that model's differences marked: components it replaces, adds or doesn't use, and how it configures the ones it keeps, with sources. Profiled so far: DeepSeek-V4.1-Flash, Inkling, Kimi K3, Nemotron 3 Ultra, DeepSeek-V4-Pro, Gemma 4 31B, MiniMax-M2.7, GLM-5, Qwen3.5-397B-A17B, gpt-oss-120b.

The Frontier LLM. A composite of the open-weight frontier: the token mixer is now chosen for KV bytes per token (hybrid linear attention, compressed or sparse softmax attention, or local/global interleaving), the residual stream became a design surface again (mHC, Attention Residuals), MoE got sparser (16 of 896 experts in Kimi K3), and Muon displaced AdamW for matrix weights at trillion-parameter scale.

The Frontier Agent. A single-threaded loop with most of the engineering in the control plane: context moved into files (skills, on-demand tools, server-side compaction), autonomy is gated by classifiers inside sandboxes, harnesses and sandboxes became disposable around a durable session log, and memory gained offline consolidation.

Causal encoder-decoder MoE: decoder reads encoder-projected KV, CSA2 shares KV across layers, FP4 KV at 890 B/token, Engram, DSpark.

5:1 sliding-window/global GQA with a learned relative-position bias instead of RoPE, short causal convs, encoder-free image and audio.

93 layers of 3:1 KDA / gated MLA with no positional encoding, Block Attention Residuals and a 16-of-896 latent MoE in MXFP4.

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.

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.

Dense 5:1 local/global stack: 1,024-token windows, global layers with K = V and p-RoPE, sandwich norms, tied 262K vocabulary.

Full GQA attention in all 62 layers after hybrid-attention ablations; 256-expert top-8 sigmoid MoE with 9.8B active; 3 MTP modules.

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.

15 × (3 Gated DeltaNet + 1 gated GQA) layers over 512 small experts: only 15 of 60 layers keep a KV cache.

A deliberately plain MoE: alternating 128-token banded and dense GQA layers, attention sinks, MXFP4 experts.