The number worth carrying out of the paper trail is a ratio. DeepSeek-VL's final pretraining mixture stayed at roughly 70 percent text against 30 percent multimodal data [9], which is about 2.3 parts text for every part of image-bearing data [10]. The schedule around it points the same way. The adaptor trains first while the vision and language components stay frozen; then the language model and adaptor train together on text-only and multimodal data with the main vision encoder still frozen; then supervised fine-tuning runs on multimodal instructions and text conversations [8]. Multimodal data is phased in rather than switched on [11].
What the encoder is for is legible in its parts. A lower-resolution semantic branch based on SigLIP-L runs alongside a higher-resolution branch derived from a SAM-B-style encoder, on the stated grounds that global semantic understanding does not resolve small text, dense documents, OCR or visual grounding [7]. Above it sit a vision-language adaptor and a DeepSeek language model [6]. The instruction data came from a taxonomy of real user scenarios, covering transcription, conversion, multi-image comparison and safety-related prompts among other tasks [5].
VL2 is where the engineering intent stops being ambiguous. Its language component is a Mixture-of-Experts paired with DeepSeek's efficient attention design, so that added multimodal capability does not make every parameter active for every token [16], and it shipped as Tiny, Small and larger variants at the point where the stack is built around high-resolution inputs, structured documents and cheaper inference rather than general image conversation [18].
Janus complicates any reading of the word "vision" in a product name. Its argument is that understanding and generation need different visual granularities, and that a representation that is excellent at recognising a chart is not automatically the right one for generating pixels [20]. An endpoint that accepts images inherits one side of that argument and tells you nothing about the other.
The inspectable portion of all this is narrower than the lineage suggests. DeepSeek released roughly 1.3B and 7B VL variants with project code and weights [13], which is more than a closed endpoint gives you and is also two research generations behind the model now being discussed. The published record explains, in some detail, what the flash vision model was probably designed to be good at. It does not document what it was made from.
