DeepSeek Efficiency Gains Expose Limits of US Chip Controls

DeepSeek's open releases demonstrate mixture-of-experts designs that achieve frontier-level results with far lower active parameter counts than dense US models.

Twitter thread notes cite DeepSeek-V2's 236B total parameters activating only 21B, matching Llama-3-70B on key benchmarks per the model's arXiv technical report. This matches patterns in prior Chinese open models such as Alibaba's Qwen2 series, which similarly prioritized sparse activation over scale.

Coverage omits explicit ties to Bureau of Industry and Security export rules that restrict NVIDIA H100 shipments, yet DeepSeek GitHub training logs confirm use of domestic Ascend clusters. Related Hugging Face evaluations show consistent downstream gains across successive DeepSeek checkpoints despite hardware caps.

Primary sources therefore indicate algorithmic efficiency, not raw compute access, as the decisive variable in current US-China model parity.