Summary
DeepSeek has released 4.1 Flash 1017, named for its October 17 release date, and open-sourced its inference engine, Deep Infer. This AI Search video puts the model through demanding agentic tests and explains why the release matters for open models competing with closed frontier labs.
The tests run in DeepSeek’s own agentic harness, which resembles Codex or Claude Code, with effort set to max. In the first, the model spent 36 minutes building a motion-graphics explainer on how Eratosthenes measured Earth’s circumference, using Gemini TTS for voiceover and the open-source Hyperframes tool for animation. It used around 22 million tokens with a 99% cache hit rate. The presenter says the quality rivals Claude and GPT output.
Other challenges include building a procedural 3D game in Unreal Engine with a SketchFab character and Mixamo animations, and reconstructing an Airbnb property in Blender from listing photos. That last one took about an hour and 37 minutes and produced visible errors such as clipping furniture and a missing slanted pool wall, showing a gap with GPT and Opus. A “find the frog” image test examines its visual reasoning.
The video closes with benchmarks and a look at what Deep Infer reveals about running massive models at scale.
📺 Source: AI Search · Published October 23, 2026
🏷️ Format: Benchmark Test
