MODEL ATLAS / 01

Architectures I’m tracking.

A small library of models I’m curious about, opened up mechanism by mechanism—from the whole pipeline to the head, filter, state, and expert routes inside.

04 field guidesUpdated August 2026

THE LIBRARY

Pick a model to open.

This is a watchlist, not a ranking. Each guide uses one visual grammar while preserving the architecture’s actual counts and relationships.

01 / GENOMICS

Evo 2 7B

A 32-layer genomic model that alternates three Hyena filter families with periodic global attention.

Family
StripedHyena 2
Scale
7B dense
Context
Up to 1M bases
Open the architecture

02 / OPEN REASONING

gpt-oss 20B

A sparse decoder where 64 query heads share eight key/value pairs and four experts work per token.

Depth
24 layers
Scale
20.91B / 3.61B active
Context
131,072 tokens
Open the architecture

03 / EFFICIENT SCALE

DeepSeek V3

A 671B model that compresses its attention cache and activates a fine-grained expert subset.

Depth
61 layers
Scale
671B / 37B active
Cache idea
512 content + 64 position
Open the architecture

04 / HYBRID SEQUENCES

Qwen3-Next 80B-A3B

A 3:1 hybrid of recurrent linear attention and periodic full attention with ultra-sparse experts.

Depth
48 layers
Scale
80B / ~3B active
Context
262,144 native
Open the architecture