Kandinsky 6.0 on a 16 GB GPU: official recipe OOMs, only the 29B model passed the Spanish test

A developer tested Kandinsky 6.0 (5-second video with audio and voices from a text prompt, MIT license) on a rented 16 GB RTX 4080. The official per-component CPU offload recipe ran out of memory even with the distilled Lite model, as the Qwen2.5-VL text encoder alone totals 15.45 GiB. A two-phase recipe with streamed encoder layers and a tiled video VAE worked, peaking at 11.24 GiB and about 141 s per vertical clip. In a Spanish speech test, only the 29B Pro model met the WER ≤ 0.20 threshold, scoring 0.15 versus 1.06 for Lite.

• Official CPU-offload recipe OOMs on 16 GB even with the Lite model
• Qwen2.5-VL encoder alone is 15.45 GiB; 8-bit and sequential offload failed
• Two-phase recipe peaks at 11.24 GiB, ~141 s per vertical clip, up to 45.9 GiB RAM
• Spanish test: WER 0.15 for Pro (29B) vs 1.06 for Lite (3B)