Slop TVNewsLatest
Explainers

FastVideo Dreamverse Harness takes AI filmmaking fully local

Hao AI Lab's open-source harness runs the agent, the model and the whole film pipeline on one NVIDIA DGX Spark, so scripts and takes never leave the machine.

Illustration: FastVideo Dreamverse Harness takes AI filmmaking fully local
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Hao AI Lab's FastVideo team open-sourced Dreamverse Harness, an agent harness built on FastVideo that takes local video generation from single clips to complete films.
  • On a single NVIDIA DGX Spark, the whole stack runs locally: Qwen 3.8 27B in NVFP4 serves as the agent's LLM, FastH3 renders the video, and scripts, references and takes never leave the machine.
  • The harness needs two servers, a FastVideo render server and an OpenAI-compatible chat endpoint, and scales to four DGX Sparks connected directly over QSFP cables without a switch.
  • The repository is MIT-licensed and built on DeepSeek Harness.

Hao AI Lab's FastVideo team has open-sourced Dreamverse Harness, an agent harness built on FastVideo that takes local video generation from single clips to complete films. It is a workspace built around a video model rather than a model on its own: a place where a person and an agent build a film together, from script and story beats to camera design and the final cut.

The pitch is local-first. Cloud AIGC studios such as Higgsfield, Runway and TapNow can produce a full film today, but they host the entire stack, so you pay a credit markup on every generation and your customization stops at their UI. Dreamverse Harness keeps the pipeline on hardware you own, and because it is open source you can edit the agent's logic or plug in a model of your own.

What Dreamverse Harness actually is

The team describes it as a workspace, not a chat box. Instead of hiding generation behind a single prompt field, it connects an agent to a visual canvas for your assets and a timeline for your final cut, so every character, scene asset and take stays visible and editable as the film comes together.

It ships with a FastVideo server running FastH3, either on a remote cluster or locally. To add another model or an API, you write one provider plugin, or let a coding agent write it for you with the built-in dv-add-model-provider skill.

The harness is built on DeepSeek Harness, and it treats every part of the stack as a modular plugin. You can strip out features you do not need, tune the skills and context fed to the agent, or build your own canvas nodes and UI visualizations.

One box, the whole film

The headline configuration runs entirely on one NVIDIA DGX Spark, the GB10 Grace Blackwell desktop with 128 GB of unified memory. On a single Spark, Qwen 3.8 27B in NVFP4 serves as the agent's LLM, FastH3 renders the video, and scripts, references and takes never leave the machine.

That is the part that matters for anyone holding client material. Local is the default, not a requirement: a cloud LLM or a remote FastVideo server is just another entry in the same model catalog, so you can direct with a larger cloud model while video stays on your Spark, or add remote render servers beside local ones.

The speed figures are the team's own. When FastH3 first ran on a DGX Spark, a five-second clip took 243 seconds. Today, by the team's measurement, five seconds of 480p video with audio renders in about 33 seconds on one Spark and 19 seconds on two. FastVideo is tuned for the GB10 down to the kernels, with fused NVFP4 matrix multiplies, sparse video attention and an NVFP4 video decoder, and every FastH3 model (V1, V2, Trim and Omni) ships as an NVFP4 build for Blackwell.

A separate step handles the upscale. The team adapted NVlabs' Open-H3-Regenerate-2K, which regenerates a draft at 2560x1440, to the Spark. With two cabled Sparks splitting each clip, a 2K regeneration dropped from 454 to 158 seconds, the team says.

How the harness works

Dreamverse Harness runs on DeepSeek Harness's plugin system and agent loop, and everything it adds is a plugin. The plugins stack in one direction only, from the bottom up.

The project layer records the whole life of a project. Every change, whether you make it on the canvas or the agent makes it from chat, is one record in an append-only log called records.jsonl that keeps who made the change, from which view, with which inputs, and what it produced. Undo and redo move the project's position along that history and delete nothing, so an undone take stays in the log.

References carry versions too. A character starts at one version and moves to the next when you swap its photo, and each render records exactly which versions it read. When a newer version replaces one a take used, the take is flagged so you or the agent can decide whether to render it again.

Above the project layer sit the model generation layer, which holds the catalog and the providers that serve the models, the feature layer, which holds the parts of a film such as characters, locations, shot renders and the timeline, and the workflows. The story workflow that ships with the harness previews which shots would render, with a GPU estimate and every reason a render would be refused, then writes the whole chain on your approval.

The agent is not one more layer. It works across all of them, following the video-directing skill to plan a film and acting through the tools the plugins provide.

What you need to run it

Dreamverse Harness needs two servers: a FastVideo server that renders the video, and an OpenAI-compatible LLM endpoint that runs the agent. The repository lists Node.js 22.19.0 or 24 and up with pnpm 11.7.0, ffmpeg and ffprobe, one or more NVIDIA GPUs for the FastVideo server, and an OpenAI-compatible endpoint for the agent, with deepseek-v4.1 as the default model id.

The team's own fully local recipe puts Qwen 3.8 27B at about 39 GB under vLLM and FastH3 Trim at about 45 GB while rendering on the same box. One FastH3 NVFP4 model takes 40 to 55 GB, so two fit on one Spark, or one beside Qwen.

Scaling is by cable, not by cloud. Up to four DGX Sparks connect directly over QSFP cables without a switch, and the harness treats every combination the same way. Add each Spark on the Hardware settings page and the agent sees whether each device is reachable, whether it is busy, and which models it holds, then places each clip on an idle Spark and renders independent clips in parallel.

The catch

The bar is hardware. The stock FastH3 8-step configuration uses four GPUs, and the fully local one-Spark path assumes an NVIDIA GB10 machine, which is a developer desktop rather than a laptop. The gains the team reports are its own measurements on that hardware, not independent benchmarks.

The project is also brand new. The GitHub repository showed fewer than 50 stars when it was checked on November 20, 2026, and its most recent commits are from the past few days.

What it means for creators

If the local stack holds up outside the team's own setup, the shift that matters most is in how the work is metered. Cloud studios charge a credit markup on every generation, and their customization stops at the UI. A harness that runs the agent, the model and the project history on hardware you own moves that cost from per-generation credits toward electricity, and puts the agent's logic back in your hands.

The counterweight is that local is still a smaller box. A single Spark holds the agent's LLM and a FastH3 model together, or two FastH3 models side by side, and the agent switches between them as the film needs them. Four Sparks render four shots in parallel before all four split each 2K upscale.

Dreamverse Harness is MIT-licensed. Clone it from github.com/fastvideo-project/dreamverse-harness and follow the DGX Spark guide in the repository, or run it against a cloud LLM and a remote FastVideo server.

Sources

  1. haoailab.com - the launch post; the local stack, plugin layers, DGX Spark recipes and speed figures
  2. github.com - the repository; MIT license, DeepSeek Harness base, setup requirements
  3. x.com - the team's announcement post