DeepSeek analysis

Thesis

DeepSeek is shipping open-weight models at scale while simultaneously open-sourcing systems infrastructure alongside them. Its public footprint is split between flagship reasoning/general models (the R1 and V3 lines, both ~684.5B-param checkpoints) and a steady drumbeat of systems/infrastructure releases (DeepGEMM, DeepEP, FlashMLA, 3FS) — a model-plus-infra strategy.

Shipping

Flagship weights dominate downloads on Hugging Face. `DeepSeek-R1-0528` leads at 6.44M 30-day downloads (684.5B params), with the original `DeepSeek-R1` at 5.69M downloads and the highest like count on the slate (13,376). The general-model line continues with `DeepSeek-V3.2` at 4.12M downloads (685.4B params) and the earlier `DeepSeek-V3` at 1.07M; an experimental `DeepSeek-V3.2-Exp` is also published (166K downloads).

A second product axis is OCR/vision: `DeepSeek-OCR` (2.70M downloads, 3.34B params) and its successor `DeepSeek-OCR-2` (1.84M), plus the smaller VLM `deepseek-vl2-tiny` (843K).

The R1 reasoning behavior is widely propagated through distillations into other base models: `R1-Distill-Qwen-1.5B` (772K), `R1-Distill-Qwen-32B` (554K), `R1-Distill-Qwen-14B` (547K), `R1-Distill-Qwen-7B` (531K), `R1-Distill-Llama-8B` (486K), and `R1-0528-Qwen3-8B` (291K).

On GitHub, the model repos anchor the top of the stars chart — `DeepSeek-V3` (103.7k stars) and `DeepSeek-R1` (92.0k) — followed by `awesome-deepseek-integration` (37.8k), `DeepSeek-Coder` (23.6k), and `DeepSeek-OCR` (23.3k). The infra stack is its own cluster: `Janus` (17.7k), `FlashMLA` (12.7k), `3FS` (10.0k), `DeepEP` (9.7k), `open-infra-index` (8.0k), `DeepGEMM` (7.4k), and `DeepSeek-LLM` (7.0k). Recent tagged releases skew toward infra cadence: multiple DeepGEMM builds (a string of nv_dev_* dev tags plus v2.1.1.post3) and DeepEP `v1.2.1`, alongside the v1.0.0 tags for DeepSeek-V3 and DeepSeek-R1.

Research themes

No first-party writing captured yet.

Hiring & scaling

No careers data captured yet.

Traction highlights

Hacker News attention is concentrated on a single launch: `DeepSeek-OCR` drew 1,003 points / 244 comments, an order of magnitude above the next thread, `DeepSeek-V3.2-Exp` at 309 points / 50 comments. A long tail of smaller/newer repos also surfaced — `LPLB` (43 pts), `DeepSeek-Math-V2` (9 pts), `DeepSeek-OCR-2` (7 pts), `Engram` (3 pts), and `TileKernels` (2 pts). On adoption, the most-downloaded model is `DeepSeek-R1-0528` (6.44M/30d) and the most-starred repo is `DeepSeek-V3` (103.7k stars).