Open Source Week (FlashMLA, DeepEP, DeepGEMM, DualPipe, 3FS)
Six days of open-sourced production infrastructure covering MLA decoding kernels, MoE all-to-all, FP8 GEMM, pipeline parallelism and a file system, plus V3/R1 serving numbers.
FlashMLA (Hopper MLA decoding), DeepEP (expert-parallel comms), DeepGEMM (FP8 GEMM), DualPipe/EPLB (parallelism and load balancing), 3FS and Smallpond (storage). Day 6 disclosed inference stats and a 545% theoretical profit margin.
- Date
- Monday, 24 February 2025
- Lab
- DeepSeek
- Kind
- infra
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Theoretical cost-profit margin (V3/R1 serving) | 545% 73.7k input and 14.8k output tokens/s per H800 node; theoretical, assumes all usage billed at R1 prices | company |
Open-infra repos are released under permissive licenses (the index page reports CC0-1.0). Day 1 was 2025-02-24; day 6 (inference system overview) came at the end of the week (about 2025-03-01). The 545% margin is DeepSeek's own theoretical figure, not audited.
Sources
This record was checked against its sources on 6 October 2026. How we check