MiniMax and DeepSeek Reportedly Operate Below 100,000 H100-Equivalent GPUs

China’s AI labs MiniMax and DeepSeek are reported to have compute capacity equivalent to fewer than 100,000 H100 GPUs, yet their latest models continue to perform near the frontier, according to Wccftech. The source says the compute estimate does not explain their model progress.

DeepSeek-V4.1-Flash-1017’s recent position at the top of an agentic benchmark score-versus-cost curve was short-lived. Wccftech reports that MiniMax M3.1 Flash has since taken the lead, describing the curve as a balance between benchmark performance and cost.

MiniMax M3.1 Flash has 428 billion total parameters, of which 23 billion are active, the source reports. It does not provide a price in the available excerpt.

The article says Anthropic and OpenAI have tightened restrictions on distillation attempts. It remains uncertain whether Chinese AI labs can sustain their current trajectory, particularly given their comparatively small compute capacity. The excerpt does not establish what accounts for the models’ performance.