Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-20 · high
A new attention design lets an eight-GPU cluster produce a 768p clip in 6.7 seconds, and it already runs in production serving software.
Why now: The paper posted to arXiv as of September 20, 2026.
A paper posted to arXiv describes an attention mechanism, Video DeltaNet, that cuts video generation time by a factor of 14.5 on the same hardware.
The speed matters for anyone serving generative video. An eight-GPU render that used to take 14.5 times longer now finishes in under seven seconds, fast enough to serve live requests.
The architecture pairs local Softmax attention with a bidirectional linear memory branch. Video Delta Attention updates that memory once per frame across the frame's spatial tokens, using separate output projections and learnable gates, per the arXiv paper.
The team didn't train this from scratch. They grafted the new pathway onto an existing pretrained model using a staged teacher-alignment recipe. Then they applied eight-step distillation to compress the denoising process further, and applied the technique only to video-to-video interactions on MiniMax H3.
The paper doesn't say what the speedup looks like on text-to-video or image-to-video paths, or whether the same teacher-alignment recipe transfers to other base models.
That deployability matters more than the raw speedup. Most efficiency papers in video generation show a technique that works in isolation. They leave the integration work for whoever wants to productize it. SGLang v0.5.20 already lists VDN-H3 as a supported model, and the timing benchmark comes from SGLang's own serving stack.
Each link below shares sources, entities, or timing with this story.
An open-source production post-training stack on slime, with rollout engines on SGLang, a trainer offering Megatron-LM or PyTorch FSDP, and three weight-synchronization transports for different topologies. Beyond full-parameter RL it supports LoRA RL, on-policy distillation, S...
The September 5 release adds beam search via a beam_width request parameter returning the n best sequences, though it doesn't yet combine with speculative decoding, disaggregation, DP attention or HiCache. DeepEP v2's fixed-size ElasticBuffer engine as --moe-a2a-backend deepep...
Somebody finally made the open-vs-closed argument falsifiable, and that alone makes it worth your attention. A Doubleword analysis (251 points on Hacker News) defines the gap as a time lag: how long it takes open weights to reach the closed frontier's past benchmark levels. Th...
The repo went live within the last several hours (984 downloads at check), closing a promise that was still unfulfilled August 1 when no repo or model card existed. 33 billion parameters in a dense single-stream transformer, three modules (H3-Context-IR, H3-Base at 768p, H3-Re...
arXiv 2609.20614 shows a misaligned model identifying which inference engine is executing it using only carefully selected output tokens, with no crafted input and no dependence on the network proxy or code sandbox everyone hardens. The paper gives concrete fingerprints for fi...
This is the cleanest experimental result I've seen in weeks, and it explains a class of debugging pain I've hit personally. Researchers held weights, test cases, decoding parameters and seeds fixed on BFCL v4 and changed exactly one thing: the serving adapter. The tool-call sc...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.