There are more and more DeepSeek papers. If I remember only model names, they soon get mixed up. This guide organizes the reading around a few questions: how architecture saves resources, how reasoning is trained, and how visual inputs are integrated.
Separate the main research directions
| Direction | Start with | What to examine |
|---|---|---|
| Architecture and training | DeepSeek-V2, V3 | MoE, attention, and training and inference costs |
| Reasoning and post-training | DeepSeek-R1 | Reinforcement learning, cold-start data, and distillation |
| General vision-language modeling | DeepSeek-VL2 | Image encoding, multimodal understanding, and dynamic resolution |
| Document recognition | DeepSeek-OCR, OCR 2 | Visual compression and recovery of text and layout |
These are not simply different sizes of one model. Visual question answering and PDF recognition solve different problems and need different evaluations.
DeepSeek-V2: start with the architecture
V2 helps explain two ideas: why MoE separates total parameters from parameters activated per computation, and how MLA addresses the cost of attention’s KV cache.
Sketch one layer first, then examine expert selection, shared experts, and attention separately. Fewer active parameters does not automatically mean the whole model fits in less GPU memory: weight storage and computation are distinct costs.
Start here: V2 paper、official repository。
DeepSeek-V3: connecting architecture with engineering
V3 continues the MLA and MoE approach, while discussing auxiliary-loss-free load balancing, FP8 training, and multi-token prediction. Compare it with V2 and distinguish architectural changes from training-engineering cost reductions.
Ask what constraints motivated each tradeoff. A training-cost figure has a specific accounting scope; it is neither total research-and-development cost nor a direct estimate of personal-computer deployment requirements.
Start here: V3 technical report、official repository。
DeepSeek-R1: how reasoning is trained
Read R1-Zero and R1 separately. R1-Zero explores developing reasoning directly through reinforcement learning. R1 adds cold-start data and further training stages to address issues such as readability. Track where the data comes from, what rewards measure, and which problems each stage addresses.
R1’s distilled models are not simply compressed copies of the full R1 weights. They are other models trained using generated data. Consult each model card for the base model and release.The R1 paper and official repository describe these details.
My earlier R1-32B deployment notes provide a deployment starting point, but keep the full model and distilled versions separate when reading the research.
DeepSeek-VL2 and OCR: distinguish the visual tasks
VL2 focuses on vision-language understanding. Follow its image encoder, connector, and language model. It also uses MoE, and image resolution affects input-token cost. See the VL2 paper。
For OCR, ask how a document page becomes visual tokens, whether text can be recovered, and whether tables and layout relationships survive.DeepSeek-OCR discusses visual compression of textual context; OCR 2 further examines reading order in visual encoding. The implementations are in the OCR repository and OCR 2 repository。
Practical setup and verification are covered in running models locally; this entry focuses on relationships between papers.
How I take reading notes
For each paper, record five things: the problem, core method, training data and setup, scope of experimental comparisons, and remaining limitations.
For deployment comparisons, record the exact model version, quantization, input length, and image resolution. Without those, “fast” or “good” is difficult to compare.
This is a starting-point index rather than a complete DeepSeek bibliography. Upstream results are not presented as my own reproduction results. Organizing these research directions first makes later papers easier to place.