Nunchux AI releases an attention kernel aimed at video diffusion speed
single source· 1 articles · confidence: low · first seen 2026-09-17 00:45 UTC
What this means for you
Nothing to act on yet. No code link, licence or benchmark was published, so there is no way to check the claimed speedup or test whether it fits your serving stack. If you run video diffusion models, wait for the release and a measured comparison against your current attention implementation.
Nunchux AI has released VC-Attention, a low-bit kernel (the low-level code that runs a model's arithmetic on a given chip) for the attention step in video diffusion transformers, the models that generate video. It is training-free, so it drops into an existing model without retraining, and it runs at reduced numerical precision (quantisation: storing numbers in fewer bits to cut memory traffic). The company says it targets two faults at once: error from quantising attention values, and a slow softmax stage, the normalisation step that turns attention scores into weights. The announcement carries no speedup figures or independent evaluation.
Key facts
- ·Nunchux AI released VC-Attention, described as a training-free low-bit attention kernel for video diffusion transformers. source
- ·The kernel is said to address two problems together: value quantisation error and a slow softmax stage. source
- ·Video diffusion transformers flatten a clip into a single sequence of spatiotemporal tokens and run full self-attention at every layer. source
- ·The report dated 16 September 2026 gives no speedup figure, benchmark result, licence or code link in the material available. source
What the sources say
- MarkTechPost — Announces the kernel and explains why attention dominates video diffusion cost, with no benchmark results.
Sources
The original reporting. Follow these — they did the work.