Vidu S2 rewrites the video while it is still playing
The two stories nobody outside China is covering this morning are both about the same thing: what happens to AI video and AI serving once the model has to answer in real time.
Shengshu Tech made the live video stream editable while it plays. Vidu S2, released September 16, is two models — Avatar for live interactive digital humans and Editing for changing a running video — and the avatar's live output moved from 540p to 720p while the frame rate holds, so a viewer's request to change someone's outfit, swap the person, replace the background or restyle the whole scene lands inside the stream instead of in a re-render queue. Reference images can be handed over during a conversation and take effect from a chosen timestamp, which is what makes the thing feel directed rather than prompted. The engineering, published alongside the release, is the interesting part: a Backbone-Refiner split that generates structure at low resolution and repairs detail in an asynchronous second pass, a Self-Replay Forcing training step that teaches the model to absorb its own drift instead of compounding it, and a vision-language agent that watches what has already been generated before deciding how to execute the next instruction. Shengshu also demonstrated real-time spatial video for headset viewing — the same territory ByteDance is reportedly building toward — and the version gap is 69 days from S1, which tells you what the video-model race looks like from inside.
Intel is making the argument that the KV cache belongs on the CPU, and its numbers are hard to dismiss. A KV cache is the note-taking layer of transformer inference: the Key and Value tensors for tokens already processed, kept so the model doesn't recompute its own context. It is also pure HBM appetite: for one mid-size open model's layer configuration, caching a hypothetical one-million-token context works out to roughly 147 GB of Key-Value data before the model weights take their share. Intel's answer is tiered — hot cache stays in GPU memory, cooling cache drops to CPU-side DDR, cold cache sinks to SSD, with the processor acting as traffic cop. Its KV Shrink layer pairs that scheduling with QuickAssist compression, which Intel reports lifted lossless space savings from about 10 percent to the 20–30 percent range by reordering the data before compressing. In Intel's own test — dual Xeon Gold 6554S, two Nvidia L20 GPUs, Qwen3-32B, an 80 percent cache hit rate — time to first token came in up to about five times faster than a stock vLLM baseline without offloading, with hardware compression running roughly twice as fast as software compression and adding under 10 percent of extra first-token latency. Treat those as vendor numbers on a curated setup, not a benchmark result: the honest read is that once an agent is waiting on a tool call, its session history is squatting on expensive accelerator memory, and that is a bill someone will keep trying to move.
ByteDance refreshed Doubao-Seed-2.1 Pro to a v0915 snapshot, and the changelog is all agent reliability. The update, live on the Volcano Engine Ark platform, targets long-horizon task delivery and multimodal coding rather than headline benchmark scores. It lands on top of the launch version introduced at the June Force conference, where Volcano Engine president Tan Dai framed the model as crossing a "production-grade" threshold on code delivery, long-range agent tasks, multimodal understanding and enterprise stability. A snapshot refresh about twelve weeks after a flagship's debut, published without a launch event, is how frontier labs operate; Chinese model vendors are now matching it in public.
Real-time means the model can no longer take an hour to think — does that push video generation ahead of text in the race for inference investment? Tell us in the comments.
