In early October 2026 I shared a series of X (formerly Twitter) posts about recent advances I was working on: KV‑caching for flow‑type diffusion models, speed‑memory benchmarks for Tensor‑parallel loading in 🤗 Diffusers, and the release of the MIT‑licensed Kandinsky video‑audio model.
- KV‑caching was integrated into Flux.2‑Klein‑KV for text projections (first model to support it for reference‑image tokens).\n- Combining the traditional TaylorSeer caching technique with KV‑caching led to "noticeable quality degradation".\n- GQA (grouped query attention) already works pretty well in Krea 2 series models, offering a low‑risk entry point for diffusion models.
- Tensor‑parallel loading (TP degree 4) for Flux.2‑Dev DiT on A10G GPUs reduced loading time from 30.4 s → 12.5 s and peak CPU memory per rank from 64.1 GB → 6.8 GB.\n- This yields ~2.4× faster loading and ~89 % less CPU memory.
- The new Kandinsky video‑audio model was released with 6 variants and 2 super‑resolution checkpoints, all open‑weight and integrated into 🤗 Diffusers day‑0.
Community feedback was solicited on all these releases, with thanks to contributors such as @Kun11664638 and @Jhuaplin. I also noted unrelated chatter about Valentine’s gifts, Apple Pay in India, and repo naming.