SolarWM-H3: 1.15 seconds from keypress to screen, on eight Blackwell GPUs

Project reports: its browser demo reaches a 1.15-second median for a matching frame after a keypress. The integrated backend reports 24.38 steady-state FPS. The documented route is video-only and needs eight Blackwell GPUs; no p95, long-session quality or complete bill is published. Read the browser and hardware check.

TaoMate-H3: merged WebSocket video

Project says: a vLLM-Omni server streams video with prompt updates between windows on H200-class multi-GPU profiles. It is a documented integration, not a pool-observed live deployment; merged-main MP4 has no AAC. Read the status check.

fal Worlds: hosted interactive sessions

Provider says: Abot World 0 and Lingbot World 2 can be entered through a browser and controlled through WebRTC data messages. These are available route descriptions, not an uptime or quality check. See price and use limits.

StreamRender-H3: browser integration, not a measured player delay

Project says: a code-owned racing world drives an H3 visual renderer on eight GB200s. The 218 ms warm render step, 295.58 ms GPU RGB interval and about 1.03-second browser publication intervals are different clocks. Control-to-visible delay, another game, and cost remain untested here. Check the route.

YoLive: a collective next-scene product

Provider says: YoLive lets viewers propose and vote on the next scene. Yoroll’s launch announcement says its H3 Superfast model generates 10 seconds of 768p, 24 fps video with native audio in four seconds on eight B200 GPUs.

Unknown: direct observation of a vote cycle, vote-to-screen time, moderation behavior, scene persistence, public uptime, audience activity, cost, and API availability. This is a current product-shape reference, not an independent service review.

FastH3 Live: a consumer-GPU local-channel claim

Community project says: its dataset card documents a local FastH3 channel on one Windows 11 RTX 5090 (32 GB), with clips produced while prior clips play and current source-stated 22.1-fps throughput at 448×448. It uses a self-contained scene library and retimed playout.

Unknown: independent replication, native-24-fps performance, long-run uptime, public delivery behavior, operating cost, prompt-to-play delay, moderation, and whether viewers accept its motion and quality compromises. The entry is a reproducible-format lead, not proof that a local GPU has solved interactive TV.

fal.live

Provider says: fal.live presents AI television channels and links to H3. The Director product page says each channel uses one Director session and viewers can influence the next direction through a voting flow.

Unknown: public uptime, actual audience size, moderation outcomes, restart behavior, and total operating cost. The entry is valuable as a product-shape reference, not as proof that a small operator can reproduce it.

Reactor FastH3 reference

Project report: the infinite-livestream repository exposes a chat-to-queue-to-WebRTC-to-RTMP architecture. Its client can keep a paced broadcast moving, while the FastH3 model notes say independently generated clips hard-cut with no continuity of subject, framing, or voice.

The repository is evidence of a documented implementation shape. It is not evidence of independently verified public uptime, seamless narrative continuity, or its economics.

Dreamverse FastH3: a documented frame-carried reference application

Project says: FastVideo’s current README exposes an opt-in FastH3 Preview v1 profile: four visible GPUs by default, 124-frame 768×1344 audio-video segments, and last-frame-to-next-segment conditioning. It also exposes separate health and readiness states and warns that default cold compile warm-up can take tens of minutes.

Unknown: a generated session observed by this pool; warm segment rate; ready-buffer margin; first-frame handoff fidelity; viewer-visible delay; recovery; cost; and public uptime. This is a source-backed product and deployment reference, not a tested public service.

What this index will require

A future entry needs a source, check date, route type, control contract, input path, generation or session shape, delivery path, source and delivered fps, and a clear note about what is observed versus assumed. A pretty loop alone is not a deployment record.