A short reading of what the open-source tree actually ships — synthesized from the README and public engineer explainers. Professional notes only; no flamebait.
Shipped
- Offline inference for DeepSeek-V4.1-Flash
- SM100 (CUDA 13) and Ascend 950 (CANN 9.2) — Ascend day-0 in the listed matrix
- Kernels (FlashMLA-family, DeepGEMM, DeepEP, …), EP/DP/CP, DSpark, SWA replay, structured outputs
Not shipped
- KV cache system · full PD-disagg serving · DP load balancing
- Official Docker image (community Dockerfiles = examples)
- 910C / SM90 / H20 as supported SKUs (draft RFCs only)
Production throughput and $/GPU-hour figures belong to the production system write-up — see FAQ attribution and request pipeline.