Summary
Developer and YouTuber sentdex puts two new fast models head to head: GLM 5.3 Flash and DeepSeek V4.1 Flash. The video walks through benchmark results, including an NVFP4-quantized GLM 5.3 Flash that ties for the top spot with DeepSeek V4.1 Flash, and argues that time to a completed result matters more than raw token speed or a single benchmark score.
He also explains why he prefers GLM 5.3 Flash in practice, showing side-by-side design iteration where one model fixes issues automatically while the other keeps redoing work. New tests use turn-based strategy games, with the models playing against each other in a ship-and-planet resource game.
The final section looks at decision models, including a “semantic if” approach, an open implementation of the idea behind decision-style models, for converting an ordinary model into one that makes structured choices. Viewers get a candid look at benchmark caveats, run-to-run variance, and how different models respond to this conversion.
📺 Source: sentdex · Published September 23, 2026
🏷️ Format: Benchmark Test
