Visual input settings
Image / video patches
A 3D convolution groups two temporal frames with each 16×16 spatial patch and projects RGB input to 1,152-dimensional vision tokens.
Qwen3.6-35B-A3B · checkpoint boundary
Where vision enters the language model
The Atlas now models both sides of Qwen3.6's multimodal boundary: a 27-block vision encoder produces 2,048-dimensional visual features that enter the language-model embedding stream before the 40-layer decoder stack.
Vision depth27released vision blocks
Vision hidden1,152d16 heads · 4,304d MLP
Vision output2,048dmatches language hidden width
Language depth40released decoder blocks
Checkpoint graph
The multimodal handoff
Exact configuration facts are shown as dimensions and counts. The spatial layout explains data flow; it is not a recorded activation trace.
Vision tower
27 encoder blocks · 1,152d
Every block uses pre-norm residual structure with 16-head non-causal vision self-attention and a 1,152→4,304→1,152 vision MLP. Open the vision-tower explorer for the implementation-level path.
Representation handoff
Visual features → embedding stream
Groups of four 1,152d vision states are merged through a 4,608d intermediate representation and projected to 2,048d, the same hidden width used by the language tower.
Language tower
40 decoder blocks · 2,048d
Ten repetitions of 3 × Gated DeltaNet + 1 × Gated Attention. Every language block uses the checkpoint's 256-expert sparse MoE path.
Evidence boundary
What this page claims
The Atlas keeps checkpoint truth, explanatory geometry and simulated activity as separate evidence levels.
Released configuration
27 vision blocks, 1,152d hidden size, 16 heads, 4,304d MLP, patch size 16, temporal patch 2, spatial merge 2, 2,048d vision output and 40 language blocks.
Subsystem boundary
Conv3D patch embedding, learned spatial positions, vision RoPE, pre-norm attention/MLP blocks, patch merger and visual-feature insertion into the language embedding stream are represented from released implementation evidence.
Runtime activity
The Atlas does not claim captured pixel activations, learned attention maps or real token-level activation magnitudes. On-screen motion and intensity remain explanatory.
Primary evidence