DeepSeek-V3 brings an efficiency-first open model to the frontier race
DeepSeek releases a 671B MoE model with 37B active parameters, open weights, low API prices, and unusually detailed efficiency disclosures.
What happened
Event details
DeepSeek-V3 scaled the lab's MoE and latent-attention architecture to 671 billion total parameters while activating about 37 billion per token. The report disclosed 2.788 million H800 GPU hours for the final training run and converted that compute to about $5.576 million at an assumed $2 per GPU hour. That figure covers the reported run, not the company's full R&D bill, earlier experiments, salaries, or cluster capital. The release was open-weight rather than a complete open-source training stack.
Assessment
Why it matters
V3 showed that an open-weight lab could approach frontier performance with a sharply different cost and architecture narrative, setting the stage for R1.