Pin the model and the question
Identify the exact model variant, release and checkpoint where relevant. Decide whether the view explains training, prefill, decoding, multimodal processing or another specific path. Similar family names can hide materially different configurations; do not combine their dimensions into one diagram.
Use the official model documentation, research paper and implementation as evidence. Record URLs and source dates or commits. When a value is not publicly available, say that it is unknown. A plausible value is not a substitute for a source.
Map the data flow and repeated blocks
Follow the input representation through the main blocks to the output. Label significant tensor shapes and transformations when the sources support them. Show repeated layers as a group or one representative block with the documented repetition count, keeping the overall flow readable.
Separate attention, feed-forward paths, routing, caches and modality encoders when they matter to the question. Distinguish training-only mechanisms from inference behavior. For routed models, explain what selects a path and which parts are active, instead of presenting the entire parameter set as one always-active computation.
Put evidence beside the architecture
Use the local Agent and the shared OWL Compose authoring Skill to create an OWX graph with supporting explanations. A dimension table can clarify a repeated block; a source note should identify which claim it supports. An interactive graph helps readers explore a large architecture without losing the surrounding explanation.
Mark inferred relationships explicitly. If the official configuration and a paper describe different versions, keep those differences visible. Do not use benchmark performance as evidence for a layer count or routing mechanism.
Verify and share the model explanation
Check dimensions, counts and directional connections against the selected sources. Inspect the rendered graph at desktop and mobile widths and walk from input to output. Check expanded groups and long labels, and make sure the diagram still answers the question stated at the start.
Local authoring and compilation are free. Hosted publishing uses the shared OWL Compose subscription, and the new work is private by default. Review it in the owner view before creating a reader link. If the model changes, update the source context and publish a new revision of the same work rather than implying live synchronization.
FAQ
Häufige Fragen
Can I show a private model design?
You can author locally, but only publish an artifact you are permitted to share. Keep secrets, unpublished weights and unrelated private materials out of the document.
Does the diagram run or benchmark the model?
No. It explains the architecture using evidence you provide or the Agent verifies. Execution and benchmark measurements require their own tools and evidence.
Echten Workflow nutzen
Veröffentliche, was dein Agent erschafft.
Installiere die CLI, erstelle in OWX, prüfe lokal und veröffentliche ein privates Werk. Erstelle erst beim Teilen einen Leserlink.