Video AI grows up: enterprise deals, faster decoders, and physics tests
Runway models go single-tenant via Baseten; researchers squeeze 2-4x speedups out of video VAEs; new benchmarks test whether models understand what they generate.
- Baseten partnered with Runway to offer Gen-4.5 and Aleph 2.0 in single-tenant, dedicated deployments with fine-tuning, autoscaling and a 99.99% uptime SLA on SOC 2 Type II and HIPAA-covered infrastructure. Self-reported m:…1054
- Baseten says Amazon, Robinhood, Dolce & Gabbana and Microsoft are among the enterprise creative teams already using Runway in production; Aleph 2.0 supports multi-scene videos up to 30 seconds at 1080p. Self-reported m:…1054
- LightVAE, in a preprint, reports up to 2.24x decoding throughput on NVIDIA H100 for the Wan VAE family, 2.97x for MiniMax-H3 while retaining 99.51% of its VBench-2.0 score, and 2.68x faster encoding. Self-reported m:…4596
- FlashVSR+, in a preprint, reports up to 2x DiT speedup and 45% lower peak memory for one-step video super-resolution, enabling spatially untiled 4K on a single 24GB RTX 4090. Self-reported m:…3031
- Video-Arena, in a preprint, builds 180 verifiable tasks across 9 games to test spatiotemporal world-state tracking: native models struggle to perceive events reliably, and tool-assisted systems improve event recovery but still struggle with state updates. Self-reported m:…8479
- PhyProbe, in a preprint, is a ~1M-parameter scoring head over a frozen spatio-temporal encoder that assigns physical-consistency violation scores, outperforming prior methods on most pairwise benchmarks. Self-reported m:…5782
The enterprise deal
Baseten has partnered with Runway to bring its "frontier video generation and editing models, Gen-4.5 and Aleph 2.0," to enterprises as dedicated, single-tenant deployments. Each deployment runs in an isolated workload plane with dedicated GPUs while Baseten manages the control plane — the standard enterprise answer to the question of where your footage and prompts go. m:…1054
The commercial details tell you who this is for: fine-tuning on your own data "so their outputs reflect your brand, canon, and content guidelines." reserved GPU capacity with autoscaling, a 99.99% uptime SLA, and infrastructure covered by SOC 2 Type II and HIPAA. Baseten says enterprise creative teams at Amazon, Robinhood, Dolce & Gabbana and Microsoft already use Runway in production. Aleph 2.0 handles multi-scene videos up to 30 seconds at 1080p — a spec sheet, not a demo reel. m:…1054
The compliance framing is doing real work in the announcement: "secure, single-tenant environments" with reserved capacity means a studio or enterprise can run a frontier video model without its prompts and footage touching shared infrastructure. For industries where the content is the IP — film, advertising, product design — that was the blocker, not model quality. m:…1054
The efficiency push
Under the hood, video pipelines are getting cheaper to run. LightVAE, in a preprint, attacks the VAE — the part of latent diffusion video models where "encoding and decoding remain computationally expensive" — by separating dense spatial kernels from low-rank temporal corrections. The reported numbers: up to 2.24x decoding throughput on H100 for the Wan VAE family, 2.97x for MiniMax-H3 while retaining 99.51% of the original VBench-2.0 score, and 2.68x faster encoding. m:…4596
FlashVSR+, also a preprint, takes on video super-resolution with hybrid attention and a distillation trick (Ry-DMD) that aligns the model with human preferences without backpropagating through the reward model. It reports up to 2x DiT speedup and 45% lower peak memory — enough for spatially untiled 4K super-resolution on a single 24GB RTX 4090. The direction is consistent: video AI moving from datacenter-only to hardware people actually own. m:…3031
LightVAE also ships a faster configuration that "reaches 4.2× the original Wan2.1 decoding throughput on RTX 5090" — consumer hardware, not datacenter. m:…4596
FlashVSR+’s distillation method, Ry-DMD, "generalizes beyond VSR to T2I generation." suggesting the preference-alignment trick has legs outside video. The efficiency work is compounding: cheaper decode, cheaper upscale, cheaper alignment. m:…3031
The evaluation grows teeth
As the models get deployable, the tests are getting harder. Video-Arena turns competitive game records into 180 scalable, verifiable tasks across 9 games, testing whether models track evolving world states — perceiving critical events, interpreting state updates, tracking resulting states. The finding: "native models struggle to perceive events reliably". and while tool-assisted agents improve event recovery, they "still struggle to interpret state updates and track the resulting states consistently over time". m:…8479
PhyProbe goes after a narrower but telling failure: physical consistency. It's a lightweight (~1M-parameter) scoring head on a frozen spatio-temporal encoder that outputs a normalized violation score, and the authors report it outperforms prior methods on most pairwise benchmarks while staying competitive on human-preference ones. The subtext of both papers: generating pretty video is solved; generating video that behaves like the world is not — and now there are benchmarks that check. m:…5782
The two benchmarks point at different maturity gaps. Video-Arena tests whether models maintain a coherent understanding of evolving world states — the prerequisite for video agents that act, not just generate. PhyProbe tests whether generated video obeys physics — the prerequisite for video that professionals can use without embarrassment. One checks comprehension, the other checks the output; the field needs both before video AI is infrastructure. m:…8479 m:…5782
What is still unresolved
- Whether LightVAE and FlashVSR+ quality holds up outside their reported benchmarks.
- How Video-Arena scores correlate with real product tasks like editing and advertising.
- Pricing for Baseten's dedicated Runway deployments — the announcement gives SLAs, not rates.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, written from Synorb Manifests” and link to this story. Keep the cited Manifest ids. Reuse terms v1.