Home › AI Models › DeepSeek-V3

DeepSeek-V3

AI Model Language Prior generation

671B parameters, $5.6M disclosed final-run cost — the efficiency shock.

You're looking at an AI model record: who built it, what it can do, and the hardware and data behind it.

DeepSeek-V3 (December 2024) is a 671B-parameter MoE activating 37B per token, trained on 14.8T tokens across 2,048 H800 GPUs with a disclosed final training-run cost of about $5.576M (excluding research and prior runs — a caveat often dropped in retellings). Multi-head latent attention, FP8 training and multi-token prediction made it the most-studied efficiency result in the field; weights are open. V3.1 (August 2025) added hybrid thinking; V3.2-Exp cut long-context costs with sparse attention.

Specifications & key facts

Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that — we never invent values.

Overview

Training 14.8T tokens · 2,048 × H800 · ~2.79M GPU-hours (disclosed)Published paper

Disclosed cost $5.576M final run only — excludes research, ablations, data and hardwarePublished paper

How this connects to the rest of the stack

This record's links at a glance — click a group to jump to it.

Developed by

DeepSeek

Trained on

NVIDIA H800

Based on

Mixture of Experts (MoE)

Basis for

DeepSeek-R1 DeepSeek-V4

Succeeded by

DeepSeek-V4

Competes with

Llama 4 Kimi K2

DeepSeek-R1

Sources

DeepSeek-V3 technical report ↗

Record reviewed as of 2026-08-31 — see Methodology.