Episode Details

The episode covers the release of DeepSeek V4 (Pro and Flash) alongside other model updates, showing that DeepSeek V4 is available free to try in instant and expert modes with a 1M-token context window, open-source licensing, and API access, while older DeepSeek API versions will be retired after July 24. The host reviews reported benchmarks versus Claude Opus, GPT 5.4, and Gemini, explains the Mixture-of-Experts architecture (1.6T/284B total parameters with far fewer active), and summarizes efficiency, attention innovations, training approach, and three reasoning modes (non-think, think high, think max). In live tests, DeepSeek’s non-think coding outputs for a landing page look dated, while deep-think improves results but is slower; the host still prefers Claude and found GPT 5.5 more impressive. The video ends by promoting an AI Profit Boardroom community and guides.