World Labs Debuts Frameboy: Real-Time Atlas Streaming at 1440p via Cloud GPU
Cloud GPU streaming reaches 15 FPS at 1440p, doubling rival resolution as the AMD deal awaits close
World Labs launched Frameboy on November 5, 2026, the first publicly accessible streaming demo of its Atlas world model. The application runs in a standard browser with no local GPU required, routing inference through cloud hardware to generate and stream AI-rendered scenes at up to 15 frames per second at 1440p resolution.
Speed Was the Core Bottleneck, Until Now
Before Frameboy, Atlas generated world-model output but could not stream it fast enough for interactive use. The company released Atlas in September 2026 and immediately began work on closing the gap between offline generation quality and real-time playback. Early testers could access a closed-beta web app called Viewfinder, but the system required substantial latency tolerance.
Frameboy changes that by disaggregating the inference pipeline. Cloud GPUs handle the heavy computation, compressed frames are transmitted over the network, and the browser renders the result locally. The end-to-end system reaches 15 FPS at 1440p under typical network conditions, which World Labs describes as sufficient for smooth navigation through generated environments.
The Architecture Behind Atlas
Atlas is an autoregressive diffusion transformer pretrained jointly on text, images, video, and 3D data. Rather than generating a full scene at once, the model produces one frame at a time, outputting both color and depth per frame. That per-frame output structure is what makes streaming viable: the system can begin transmitting before the full sequence is complete.
Three architectural decisions underpin Frameboy's performance. First, Atlas uses key-value caching across frames, which reduces the computational redundancy of rendering consecutive moments in a continuous environment. Second, World Labs implemented cache-aware routing to distribute inference across multiple cloud GPUs without regenerating cached context. Third, the model generates approximately 100,000 tokens per second at inference, a throughput figure the company says is essential for maintaining interactive latency at 1440p.
The first working real-time system World Labs shipped was RTFM, released in October 2025. RTFM ran on a single NVIDIA H100 and generated frames at interactive framerates but operated at lower resolution and required dedicated hardware. Frameboy generalizes that capability to cloud-routed inference accessible via browser.
A Resolution Advantage, but a Frame Rate Tradeoff
The competitive field for real-time world models is narrow. Google DeepMind's Genie 3, released in August 2025, generates 720p output at approximately 24 FPS. Odyssey-2 Pro, available as an API since January 2026, operates at around 22 FPS at similar resolution. Overworld Labs' Waypoint One is an open-source model launched in January 2026 that runs at 60 FPS on consumer-grade gaming GPUs. NVIDIA Cosmos 3, announced in May 2026, targets physical AI and robotics applications rather than interactive rendering.
Atlas at 1440p produces roughly four times the pixel count of 720p alternatives. The tradeoff is frame rate: 15 FPS falls below the 24 FPS threshold associated with cinematic smoothness and well below the 60 FPS standard for real-time interaction.
Atlas Enters AMD's Hardware Roadmap Before the Acquisition Has Closed
AMD announced an $8.2 billion all-stock acquisition of World Labs in September 2026. The deal has not yet closed and remains subject to regulatory review. Under the terms, Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist, reporting to Lisa Su.
Frameboy's release ahead of the acquisition close positions Atlas as a technology asset rather than a research project. The company has demonstrated a production-grade streaming pipeline that runs on commodity cloud GPUs and delivers browser-accessible output at competitive resolution. AMD's interest in the acquisition was partly predicated on integrating Atlas capabilities with its accelerator roadmap. Frameboy makes those capabilities publicly demonstrable before regulators have ruled on the deal.
What the Benchmarks Don't Show
Frameboy is a technical demonstration, not a product benchmark. The 15 FPS figure is a maximum under favorable conditions; real-world frame rates will vary with network bandwidth, cloud GPU allocation, and rendering complexity. World Labs has not released an independent evaluation of Atlas against standardized world-model benchmarks.
The resolution lead is also specific to streaming delivery rather than offline generation quality. Offline systems, including research models such as InSpatio-WorldFM, operate at higher resolutions and frame rates than any currently available streaming system, but they do so without the latency constraints that make real-time use viable.
The Next Milestone Is AMD Hardware, Not Another Demo
World Labs has indicated that the next phase of Atlas development targets AMD accelerator hardware. The company has not published a timeline, but AMD's Instinct MI350 roadmap positions next-generation accelerators for 2026 deployment. If the acquisition closes on schedule, Atlas workloads could move from general-purpose cloud GPUs onto purpose-built AMD silicon optimized for World Labs' inference architecture.
For now, Frameboy remains the accessible entry point. The demo is live at frameboy.worldlabs.ai and requires only a browser. It does not require an account, a waitlist, or AMD hardware. The question World Labs is implicitly answering with the release is whether Atlas can deliver a coherent real-time experience at resolution that matters — and the initial answer, with its caveats, is yes.