# Nunchux speeds video diffusion with value smoothing and FP8 softmax casting > Source: > Published: 2026-09-18 03:54:11+00:00 # Nunchux speeds video diffusion with value smoothing and FP8 softmax casting **Nunchux AI's Muyang Li and collaborators report up to 1.70x faster end-to-end generation without retraining, with MiniMax H3 among four models tested.** By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff) ยท Published Primary source: [MiniMax](https://x.com/MiniMax_AI/status/2100792780494889194) ## Why it matters Video generation costs increasingly depend on inference engineering beneath the model. VC-Attention shows how a young startup can turn academic kernel work into a credible platform wedge. [Muyang Li (@lmxyy1999)](https://x.com/lmxyy1999?ref=runtimewire) and his collaborators at Nunchux AI have published a low-bit attention method that attacks two stubborn costs inside video diffusion models: quantizing value activations and running softmax at higher precision. The September 14th [VC-Attention paper](https://arxiv.org/abs/2609.15810?ref=runtimewire) reports attention-kernel speedups of 1.46x to 1.59x on NVIDIA's H200 and B200 data-center GPUs, with larger gains on workstation cards. The work follows a path Li has been on since studying computer science at Shanghai Jiao Tong University. He later earned a robotics master's degree at Carnegie Mellon University and completed an MIT doctorate; his [doctoral thesis was titled "Full-Stack Algorithm-System Co-Design for Efficient Visual Generation"](https://www.eecs.mit.edu/eecs-events/doctoral-thesis-full-stack-algorithm-system-co-design-for-efficient-visual-generation/?ref=runtimewire). His research has included SVDQuant, Sparse VideoGen and Radial Attention, methods aimed at making increasingly expensive image and video models practical to run. That research became the foundation for San Francisco-based [Nunchux AI](https://www.nunchux.ai/about?ref=runtimewire), which Li started with CTO [Yujun Lin (@syn7xavier)](https://x.com/syn7xavier?ref=runtimewire), chief system architect Zhekai Zhang and chief scientist [Jun-Yan Zhu (@junyanz89)](https://x.com/junyanz89?ref=runtimewire). Lin completed his doctorate at MIT's HAN Lab and previously worked as an NVIDIA research scientist. Zhang led architecture and CUDA-kernel work on projects including SpAtten and the Nunchaku diffusion inference engine. Zhu, a Carnegie Mellon professor, previously worked at Adobe Research after completing a UC Berkeley doctorate and an MIT postdoctoral appointment. Li and Zhu framed their company around a simple constraint in a [September 3rd launch post](https://www.nunchux.ai/blog/building-the-frontier-of-multimodal-inference?ref=runtimewire): generated media has to arrive quickly enough for a creator to respond, while remaining cheap enough to support repeated attempts. "Video presses hardest on all three," they wrote, referring to latency, cost and quality. ### The arithmetic behind the speedup Video diffusion transformers turn frames into long sequences of spatial and temporal tokens, then repeatedly calculate how those tokens interact. As resolution and duration increase, attention becomes a growing share of the workload. Lower-precision arithmetic can make the matrix multiplications faster, but the VC-Attention authors argue that two surrounding operations have become the limiting factors. The first is the value side of attention. Earlier low-bit methods concentrated on smoothing queries and keys before quantization. VC-Attention's V-Smooth component instead groups similar value tokens using lightweight online clustering, subtracts each block's mean and quantizes the residual. It restores the mean during the online softmax calculation, avoiding a separate pass over the data. The second component, ExpCast-FP8, targets softmax itself. The method maps log-domain attention scores directly into FP8 probability codes using a fused multiply-add, replacing an FP32 exponential and the subsequent conversion. VC-Attention requires no model-specific retraining, which matters for an inference provider serving models from several labs: the optimization can be applied at the kernel level rather than baked into each model during training. In its September 18th post, [MiniMax](https://runtimewire.com/models/fal/minimax-preview-speech-2.5-hd) placed VC-Attention beside FlashAttention and SageAttention in what it called the "dense, done faster" branch of video optimization. That branch preserves the full attention pattern while making each interaction cheaper. MiniMax contrasted it with sparse methods that skip or approximate selected interactions and hybrid or linear architectures that change how information moves through the model. The categories can overlap. That is strategically useful for Nunchux: a cheaper dense kernel can be combined with Li's earlier work on sparse attention, allowing an inference stack to reduce both the number of interactions and the cost of the interactions it retains. ### Kernel gains are larger than clip gains The [full VC-Attention paper](https://arxiv.org/html/2609.15810v1?ref=runtimewire) tested Wan2.2, [LongCat-Video](https://runtimewire.com/models/fal/longcat), HunyuanVideo-1.5 and MiniMax H3 on 100 MovieGen Bench prompts apiece, with prompts and random seeds shared across methods. At eight-bit precision, the authors report that the attention kernel ran 1.59x faster than BF16 FlashAttention-4 on an NVIDIA B200 and 1.46x faster on an H200. At four-bit precision, V-Smooth reached 2.27x on an RTX PRO 6000 and 3.58x on an RTX 5090. Those numbers describe the attention call, not the entire generation pipeline. On Wan2.2, the [reported end-to-end gains](https://arxiv.org/html/2609.15810v1?ref=runtimewire) were 1.19x on B200, 1.13x on H200, 1.36x on RTX PRO 6000 and 1.70x on RTX 5090. The gap is a useful measure of what VC-Attention does and does not accelerate: model loading, other neural-network operators and output processing remain outside the kernel. For the MiniMax H3 benchmark, the paper reports a 1.60x attention speedup over BF16 FlashAttention-4 on B200. Its comparison with SageAttention2 also benefits from a hardware mismatch: the authors note that SageAttention2 lacked a Blackwell-specific kernel and ran code designed for earlier GPUs. The paper's larger competitive multiple should therefore be read alongside the cleaner FlashAttention-4 comparison. The fidelity results are also paper-reported rather than independent production tests. The authors measured PSNR, SSIM, LPIPS, subject consistency and imaging quality against BF16 FlashAttention-4 output from the same model and random seed. The [paper's fidelity tables](https://arxiv.org/html/2609.15810v1?ref=runtimewire) show V-Smooth leading the reported fidelity measures overall among the tested training-free low-bit methods. Across the evaluation, adding ExpCast-FP8 gave up 0.7 to 2.1 dB of PSNR in exchange for speed. On MiniMax H3, both variants recorded 20.2 dB PSNR, while the ExpCast-FP8 variant posted slightly worse SSIM and LPIPS. That scope distinguishes this result from [MiniMax H3's earlier faster-than-playback demonstration](https://runtimewire.com/article/nvidia-sol-h3-minimax-video-faster-than-playback), which used eight NVIDIA B300 GPUs, a four-step acceleration adapter, warmed hardware and excluded MP4 encoding. VC-Attention works lower in the stack, focusing on the attention kernel across several models and GPU classes. ### Nunchux is selling the layer beneath the model Nunchux came out of stealth on September 3rd with a serving platform for image, video and world models. Nunchux says its earlier SVDQuant work has accumulated four million downloads and has been integrated into software including NVIDIA TensorRT, AMD Quark, vLLM, SGLang, Diffusers and ComfyUI. Nunchux also names [Emergence Capital](https://www.emcap.com/?ref=runtimewire) and [E14 Fund](https://www.e14.vc/?ref=runtimewire) as lead investors, alongside Archerman Capital, Scale Asia Ventures, Tectonic Ventures, KungHo Fund and Plaid Matrix Fund. VC-Attention gives Li's young company a technical demonstration tied to models customers already recognize rather than another general claim about faster inference. MiniMax's participation helps because [MiniMax's current model catalog](https://platform.minimax.io/?ref=runtimewire) positions H3 as an open, general-purpose video model supporting text-to-video, image-to-video and multimodal-reference workflows. The evidence remains a research benchmark and an endorsement from a model provider. It does not establish a production deployment inside MiniMax's serving stack. For Nunchux, the immediate value is still concrete: the paper shows where its founders believe the next inference gains will come from. Model labs continue to spend heavily on larger video systems, while Li's team is betting that the companies controlling kernels, quantization and serving software can claim a meaningful share of the resulting economics.