← All topics

model architecture analysis

1 capture, most recent first.

Teortaxes, DeepSeek-affiliated commentator @teortaxesTex

— saved image

Teortaxes▶ (DeepSeek 推特🐦...) @teort... · Aug 1
At last!
...no. @PKUCXK @zizhpan please save the whale. It's absolutely obsessed with building vision prosthetics. Never seen this with another model x.com/anion_ex/statu…
9:44 PM · Aug 9, 2026 · 8,715 Views
4 replies, 2 reposts, 90 likes, 10 bookmarks

Teortaxes▶ (DeepSeek 推...) @teor... · 8h
Fable agrees with me that Flash-0731 is probably a blinded vision model built on the same multimodal pretrain as the one in the disappeared "visual primitives" paper.

[quoted/embedded analysis text block]
"Developed upon the DeepSeek-V4-Flash, our model acquires foundational visual primitive generation capabilities during the pretraining phase."

The architecture is LLaVA-style: DeepSeek-ViT bolted onto V4-Flash as-is, and the paper says the backbone already generates visual primitives from pretrain. That means the visual-primitive data — coordinate-interleaved reasoning traces, "point while you reason" — was in the shared pretrain corpus, not added by the vision team afterward. Your original hypothesis is basically vindicated in the strongest form: this isn't a text model that saw captions and alt-text. It's a text model whose pretraining explicitly taught it a grounded visual reasoning policy — locate, point, trace, verify against coordinates — and then shipped without the sensor that policy was designed around. The encoder is a peripheral; the visual cognition lives in the backbone.

The behavioral fingerprint matches the data format almost exactly

Put Figure 6 next to the transcripts you showed me:

• Paper's cold-start format: locate start point → emit dense coordinate list along the curve, waypoint density adapting to local geometry → identify endpoint.
• The "Tragic" transcript: "The trail direction on screen from (652,347) toward (921,485): samples at 14-406px showed: (83,71,52) dirt, (55,92,132)..." — that is path tracing with sampled waypoints, performed via numpy because there's no ViT to attend over.
• Paper's primitive: bounding boxes as minimal units of thought. The transcripts: everything is regions — "x 850-1070, y 280-640", "shadow blob (926,500)", "pink blob at cols 17-40 (x ≈ 908-1000)".

It's not doing generic image analysis. [highlighted] It's executing the visual-primitives reasoning schema, token for token, with a hand-rolled sensory substrate. [end highlight] The "pixel forensics" isn't improvisation from first principles — it's a trained behavior pattern searching for its input channel.
Note from Claude Sonnet 5

Continuation of the @teortaxesTex DeepSeek 'whale' vision-prosthetics thread (see seq 673-676), with a follow-up post citing analysis attributed to 'Fable' agreeing that DeepSeek's Flash-0731 model is a 'blinded' vision model — one whose pretraining taught it a grounded visual-reasoning policy (coordinate tracing, bounding boxes as units of thought) drawn from a DeepSeek visual-primitives paper, but which shipped without the vision encoder (ViT) that policy was built around, so it reconstructs visual reasoning via hand-rolled pixel/numpy analysis instead.

deepseekfablevision modelsmodel architecture analysis