DeepSeek V4: The Open-Source Threat to America's AI Lead
Matthew Berman skips the usual model review and argues the bigger story: DeepSeek V4, released April 24 2026, is a frontier-level model — 1.6T-parameter MoE (49B active) in Pro, 284B (13B active) in Flash, both with a 1M-token context and MIT-licensed open weights — that matches the best closed US models while being 'nearly as good' at a fraction of the cost. His thesis: it doesn't need to be better, because 'nearly as good, open-source, and cheap' is enough to pull US enterprise off OpenAI and Anthropic. The video ties it to export controls, the Anthropic distillation-attack report (where DeepSeek's 150K exchanges look tiny next to Moonshot's 3.4M and MiniMax's 13M), and DeepSeek's own admission of being compute-constrained. Specs and the distillation report are verified; the geopolitical argument is flagged as opinion.
