In October 2026 I posted on LinkedIn about three key developments: KV‑caching for flow‑based image generation, the MIT‑licensed Kandinsky video+audio model, and a Tensor‑Parallel loading upgrade for Flux.2‑Dev DiT in 🤗 Diffusers.
The Kandinsky video+audio release offers SIX different variants (including few‑step ones), balancing speed, memory, and quality.
It also comes with TWO super‑resolution checkpoints to provide more fidelity to the generated outputs 🤌.
On Flux.2‑Dev DiT, with a TP degree of 4 (A10G): loading time reduced from 30.4s → 12.5s.
Peak CPU memory per rank dropped from 64.1 → 6.8 GB.