Nunchux AI exits stealth: ex-Berkeley CycleGAN author Jun-Yan Zhu builds 100x-faster multimodal inference
songhan_mit · x · 2026-09-04
Jun-Yan Zhu, creator of pix2pix and CycleGAN, has launched Nunchux AI, an inference infrastructure for image and video generation models emphasizing speed and cost.
- The team traces the mission to Berkeley nights a decade ago making iGAN/pix2pix/CycleGAN real-time; image model inference now costs 10,000x more, with video two orders of magnitude beyond that.
- The site claims 100x inference speedup, 10x cost savings, 99.9% SLA, and 3,500+ developers onboard.
- Offers pay-as-you-go APIs with a library including Nano Banana Pro, FLUX.2 Klein, and Kling V3, plus customization.
- A demo contrasts timeout failures on an RTX 3090 with a 0.08s generation in Nunchux's Radical Speed mode, claiming up to $900k/yr in savings.
Related event: Nunchaku Team Launches Nunchux for Fast Generative AI Inference(3 posts)→
More from Infra
- DeepSeek's token density yields ~$38/GPU-hr revenue vs $3.50 for raw IaaS, engineer estimates — bookwormengr · 2026-10-21
- Sizing a 1 GW datacenter: ~522,000 Blackwell GPUs fit inside — bookwormengr · 2026-10-21
- At 85% billable uptime, a 1 GW fleet supplies 3.89B GPU-hours/year — bookwormengr · 2026-10-21
- DeepSeek's production density reportedly hits ~$38 per GPU-hour at saturation — bookwormengr · 2026-10-21
- PyTorch details KTIR and Dataflow Scheduler behind IBM's Spyre accelerator backend — PyTorch · 2026-10-21
- Indie dev replaces $1k OpenAI bill with a YAML pipeline on spot GPUs and vLLM — cmm324 · 2026-10-21