Runway Solaris really does generate visuals in real time. But the most important part of that statement is not “real time.” It is what Solaris generates: not a conventional video clip, but an interface that continuously redraws itself as the user clicks, drags, or types.
That distinction makes Runway Solaris a useful lens for the next phase of AI video. MiniMax H3 has made a capable multimodal foundation available with open weights, while fal’s post-trained H3 Max can produce a five-second video in roughly three seconds. Solaris points to the next step: generated video becoming a live computational medium.
What Runway Solaris Actually Is
Runway announced Solaris on August 31, 2026, calling it the first in a family of “Interface World Models.” It is built on Gen-4.5 and follows Runway’s GWM-1 research, but it has been adapted to understand user interaction and respond at interactive speed.
Instead of generating application code, Solaris synthesizes the interface itself, frame by frame, at 720p. A language model decides how the scene should evolve; the world model renders it. Runway attributes the speed to autoregressive generation, few-step distillation, and training the fast model on its own outputs for longer-term stability.
So yes, Runway Solaris is real-time video generation—but it is not currently a general-purpose “instant movie” model. It is an early-access interactive system designed for generated interfaces, adaptive tutorials, virtual storefronts, and agent-training environments.
“Real Time” Now Describes Three Different Things
The AI video industry increasingly uses one phrase for three distinct performance levels:
Level | Meaning | Current example |
|---|---|---|
Faster than playback | A complete clip renders in less time than its duration | fal H3 Max |
Streaming generation | Frames arrive continuously while the result is being viewed | Runway’s GWM direction |
Interactive generation | New user actions alter the next frames with sufficiently low latency | Runway Solaris |
H3 Max is an important breakthrough at the first level. fal reports that its post-trained MiniMax H3 variant generates a five-second, 768p video in about three seconds—roughly 35 times the throughput of the official H3 endpoint in fal’s testing. That is faster than real time as a rendering ratio. However, the user does not necessarily redirect the clip halfway through generation.
Solaris aims at the third level. Runway notes that interaction stops feeling responsive at around half a second of delay. Its system therefore generates sequentially and incorporates clicks, drags, and text input into subsequent frames. This is the difference between receiving a fast result and inhabiting a live result.
MiniMax H3 and fal H3 Max Change the Development Model
MiniMax H3 accepts text, images, video, and audio as context and can produce four-to-fifteen-second video with 32 kHz stereo audio at 24 fps. Its open H3-Base checkpoints generate at a 768-pixel short side; the hosted workflow can regenerate output up to 2K.
The release is not completely self-contained: MiniMax’s H3-Context-IR orchestration layer and H3-Regenerate-2K service remain hosted. Still, the base weights allow outside teams to tune the model and redesign the inference path.
fal’s H3 Max demonstrates why that openness matters. fal says it added post-training data, improved prompt adherence and aesthetics, and co-designed the model with its serving stack. The market may therefore reward not only foundation-model training, but also teams that can optimize and serve those models for a specific experience. Lightricks’ open, faster-than-playback LTX work points in the same direction.
What the Published Benchmarks Do—and Do Not—Show
These vendor-published results measure different tasks and should not be treated as a direct Solaris-versus-H3 ranking.
System | Published evaluation | Reported result |
|---|---|---|
Runway Solaris | 250 participants, 30 interaction examples, nearly 7,500 pairwise judgments | Preferred over a coded comparison in 61% of instruction-following judgments and 71% of natural-behavior judgments |
fal H3 Max | Head-to-head human preferences across 12 video models | fal reports first place for overall preference, prompt understanding, and aesthetics |
fal H3 Max | Generation latency | Approximately 2.4–3 seconds for a five-second video |
The Solaris study tests interactive behavior; H3 Max tests completed-video preference and speed. Neither establishes universal superiority. They reveal two frontiers advancing at once: interaction quality and production efficiency.
Pricing and Availability as of September 2, 2026
System | Access | Published price |
|---|---|---|
Runway Solaris | Partner rollout and early-access request | Not announced |
MiniMax H3 API | Available from MiniMax | $0.08 per second at 768p; $0.13 per second at 2K |
fal H3 Max | Available in fal’s playground and API | Promotional 768p rate of $0.02 per second through September 7; $0.08 afterward |
Solaris has no public production pricing, model weights, or generally available API yet. Its economics therefore remain unproven. Runway explicitly acknowledges that continuously generating every frame costs more than serving a conventional webpage, even after large efficiency gains.
The Next AI Video Race Will Be About Runtime
The first era of generative video asked whether a model could make a convincing clip. The next asks whether it can remain coherent while the world changes around it.
That shift has several consequences.
First, prompts will become event streams containing gestures, speech, camera movement, application state, and verified data.
Second, hybrid systems will win first. Solaris still struggles with stable text, long-session coherence, factual trust, accessibility, and software integration. Deterministic code can handle transactions, permissions, and screen readers while generative rendering supplies adaptive visual behavior.
Third, latency becomes a creative capability. At minutes, AI is a renderer; at seconds, a collaborator; inside the action loop, an environment.
Verdict: Runway Solaris Is a Direction, Not Yet a Replacement
Runway Solaris is genuinely real time according to Runway’s published design: it streams 720p frames and changes them in response to user input. But it should not be confused with a generally available, faster text-to-video service.
Its larger significance appears when viewed beside MiniMax H3 and fal H3 Max. Open weights make rapid specialization possible. Post-training and inference co-design make high-quality clips faster than playback. Autoregressive world models turn that speed into continuous interaction.
The likely future of AI video is therefore not simply longer, sharper movies generated more quickly. It is video becoming a runtime: a responsive layer for entertainment, software, training, commerce, robotics, and agent simulation. Solaris is an early and incomplete example—but it makes that future much easier to see.
