Independent analysis traces DeepSeek's vision lineage from DeepSeek-VL to Vision-Exp

An independent technical analysis published on dev.to argues that DeepSeek's newest experimental vision model is best read as the latest step in a multi-year research program rather than a sudden pivot to images. The piece, written by zipflow.xyz, reconstructs the company's public vision-language lineage from the original DeepSeek-VL through VL2 and the Janus image-generation models, then sets that history against the documented behaviour of the current deepseek-v4-flash-vision-exp API. The author is explicit that the analysis is unofficial and that the model is not exposed through their own upstream channel, so the API details are documentation-based rather than production-tested.

DeepSeek-VL treated documents as first-class vision inputs

According to the analysis, DeepSeek's 2024 paper “Towards Real-World Vision-Language Understanding” framed vision around practical inputs — web screenshots, PDFs, OCR, charts and knowledge-oriented content — rather than general image captioning. The project derived a task taxonomy from real user scenarios and used it to build instruction-tuning data spanning recognition, transcription, conversion, analysis, commonsense and logical reasoning, multi-image comparison and safety prompts.

The architecture combined a hybrid vision encoder, a vision-language adaptor and a DeepSeek language model. The encoder paired a lower-resolution semantic branch based on SigLIP-L with a higher-resolution branch derived from a SAM-B-style encoder, on the reasoning that global semantic features alone are inadequate for small text, dense pages and visual grounding.

Training ran in three stages: an adaptor warm-up with the vision and language components frozen, joint pretraining on mixed text and multimodal data, and supervised fine-tuning. The pretraining mixture stayed roughly 70% text and 30% multimodal, with a gradual modality warm-up rather than an abrupt switch to images. The analysis draws a broader lesson from that balance: adding an encoder is not sufficient, because training also has to preserve the base model's existing reasoning and instruction-following behaviour. DeepSeek-VL shipped roughly 1.3B and 7B variants along with code and weights, making the early line more inspectable than a closed API model, even if not every training sample was publicly traceable.

VL2 added dynamic tiling and mixture-of-experts

DeepSeek-VL2 extended the line in two ways, per the analysis. Dynamic tiling handled varied aspect ratios and higher resolutions without the detail loss that a fixed resize inflicts on document and screenshot tasks. A Mixture-of-Experts language component with DeepSeek's efficient attention design let multimodal capability grow without activating every parameter for every token. The paper's evaluations covered visual question answering, OCR, document understanding, table and chart understanding, and visual grounding, and the release came in Tiny, Small and larger variants.

Janus split understanding from generation

The Janus line addressed a different question: whether one autoregressive framework can both understand and generate images. It decoupled visual encoding into a semantic path for understanding and a discrete visual-token path for generation, both feeding a unified transformer, on the argument that the best representation for reading a chart is not the best one for drawing pixels. Janus-Pro later scaled the model and refined its training and data strategy. The analysis cautions against conflating this work with the current API offering: Janus is unified understanding-and-generation research, while Vision-Exp is an API model that accepts images alongside text. What they share is the underlying investment in visual representation.

What the Vision-Exp API documents

DeepSeek's documentation, as summarised in the analysis, lists deepseek-v4-flash-vision-exp as an experimental model accepting JPEG, PNG, GIF and WebP images via base64 data URLs, public image URLs or Files API references. It is reachable through OpenAI-compatible Chat Completions and Responses APIs as well as an Anthropic-compatible Messages endpoint. Limits include a maximum of 384 tokens per image after resizing, up to 600 images per request, and different caps for external or base64 inputs versus Files API references.

The provenance gap

The papers offer research-level provenance — targeted task categories, broad data mixtures, selected public datasets, internal scenario taxonomies and staged alignment procedures — but, the analysis stresses, not a source-level ledger. There is no public complete list of images used for the current model, no full licensing chain, no deduplication or contamination report, and no mapping from present model behaviour to specific training examples. Architecture and API documentation answer how to call the model, not where every training example came from.

Why it matters

The lineage explains the current API's shape. Because the early research explicitly targeted screenshots, PDFs, OCR and charts, image input in the new model speaks to document-heavy and agent-style workflows rather than photo captioning alone. At the same time, the analysis draws a distinction that matters for technical, legal and procurement due diligence: a visible research lineage is not the same as verifiable training-data provenance, and the newest model's data story remains undocumented.