DeepSeek-V3 Technical Report: read the paper

What does a reported training bill actually include?

This report connects sparse model design, numerical precision and systems engineering to a concrete training-compute account.

Baseline, result & limitation

Baseline / setup: The architecture builds on DeepSeek-V2. Table 1 accounts separately for pretraining, context extension and post-training.
Reported result: The authors report 2.788 million H800 GPU-hours. At their assumed $2 per GPU-hour, the rental equivalent is $5.576 million.
Keep in mind: That figure excludes prior research and ablations. It is neither the company’s total development spending nor a current cloud quote; hardware and reporting scopes matter.

Before you read A GPU-hour is one GPU used for one hour. More GPUs can shorten a run without reducing its total GPU-hours.

Check the original manuscriptIntroduction, Table 1: training costs and exclusions ↗Sections 2 and 3: architecture and infrastructure ↗

Artificial Post reading note · reviewed September 12, 2026

RESEARCH SUMMARY

A starting point.

DeepSeek-V3 describes a mixture-of-experts language model and techniques for efficient large-scale training.

What to keep in mind

Read the original paper for the experimental setting, baselines, and limitations; a result in one setting is not a guarantee of performance elsewhere.

This work is included in a researcher’s reading path. A detailed editorial explanation is still being prepared. The complete manuscript is available in Full paper.

Source: DeepSeek-V3 Technical Report. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.