Summary
Matthew Berman runs DeepSeek’s latest model, V4.1 Flash, through a set of practical tests and calls it the best open-source model available right now.
The tests start with a hyperrealistic interactive 3D mechanical watch diagram with hover labels and an exploded view, then a Fall Guys clone called Bean Blitz that he says comes close to Claude Opus 5.5 at a fraction of the price. A 3D iPhone Duo model, which requires the model to research the device’s specs and design, is compared with the same test run on the new Kimi model. A visual counting test shows the model reasoning about items hidden behind others in a pile of pastries.
Berman also measures speed. With reasoning off, a 1,000-word story streams at an estimated 125 to 150 tokens per second on his setup, and Artificial Analysis lists hosted speeds of up to 1,286 tokens per second from Lithos AI and 845 from Baseten. In a browser task, the model spent about six minutes finding the top five clippable moments in a Sam Altman interview and returned accurate timestamps, with its full chain of thought and tool calls visible thanks to the open weights. The video includes a CoreWeave sponsor segment.
📺 Source: Matthew Berman · Published October 28, 2026
🏷️ Format: Benchmark Test
