7-minute Manim-style explainer on Deepseek-v4.1-flash and KV cache

TLDR: Opus 5.5 built a small 3Blue1Brown style animation engine from scratch in python + ffmpeg to animate this video. Deepseek-v4.1-flash is a 552B model with a 1M-token context that keeps only 890 bytes of kv cache per token. Here's a 7-minute video explaining that, in Manim visual style, with subtitles and voice over. This was made with Opus 5.5, i asked it to read paper and write script, storyboard. It did these almost perfectly. Then generated video with it's animation engine. The paper also doesn't show how 890 bytes breaks down, so it rebuilt it from the config Voice over is generated by myself, with help of caption that Opus generated. Using edge-tts, an AI voiceover was made, with "en-US-AndrewNeural" as voice option. Merged with ffmpeg on my pc. It's hours and hours of task, condensed to 30min. Insane how good the model is.

— AJ (@ItsmeAjayKV) Sep 25, 2026