$ cat wiki/models/gwm-worlds-2.md
GWM Worlds 2
Spec
| Attribute | Value |
|---|---|
| Developer | Runway |
| Released | not yet |
| Announced | 2026-09-03 |
| Context window | unknown |
| Pricing | unknown |
| License | unknown |
| Availability | research preview; access by contacting Runway through its research page |
| Four rows need their reading stated | |
| (source): |
Releasedisnot yet, not a date. GWM Worlds 2 is announced as a research preview, and one pass states in as many words that no general release date has been announced. The 2026-09-03 date belongs inAnnounced.Context windowisunknownand the gap is a category mismatch. This is a streaming audio-visual model; nothing read states a token budget, and the nearest stated quantity — that generation is "not restricted to a fixed duration" — is a horizon claim, not a context window, and is not written into this row.Pricingisunknownbecause access is contact-only. No rate card, no hosted endpoint, no third-party catalogue entry for this model appears in anything read.Licenseisunknown— no weights, no licence text, and no statement about either.
Release Date
Announced 2026-09-03, captured here 2026-09-26 — day +23 (source).
The lateness is a finding about this pipeline, not about the release.
Runway is not in sources.yaml in any tier — not a blog, not an
X account, not in the Chinese-lab rotation — so nothing polled it and nothing
reported it missing. It surfaced only because Latent Space published an
analysis piece 22 days later (prefetch candidate #34, 2026-09-25) and that feed
is tracked. A release reaches this wiki through a commentator or not at all.
This is the same class as the Anthropic Alignment Science gap recorded
2026-07-31 — a source can be cited and never fetched — with one difference that
makes it worse: Alignment Science was at least listed. sources.yaml is
manual by this repo's automation boundary, so this run does not edit it.
Not read first-party. runway.com answers EGRESS_BLOCKED, so the research
post is unread; Runway's own X text is quoted second-hand through two passes.
Benchmarks
None. Not one figure, for this model or any comparator.
Nothing read carries a benchmark, a human preference study, a physics- or temporal-consistency score, or a comparison against any other world model. The demonstration is the evidence.
Runway's own stated caveat is that real-time generation "still trades fidelity for speed" (1 pass) — an admission of a trade-off with no figure on either side of it.
This sits against 2609.28654, captured in the same run, which built a 300-question exam and ran 14 video models through a blind pairwise Elo study. An instrument for this branch exists; this release was not put through one. See World Models.
Use Cases
The stated interface is WorldPrompt, described in two passes as a prompting mechanism, not a scripting language: it fixes some aspects of a simulated environment — including the first frame — and then accepts a series of timestamped events steering characters, cameras and environments. Events can be prompted in real time while generation runs.
Three stated modes:
| Mode | Stated use |
|---|---|
| ahead-of-time | filmmaking, directing |
| turn-based | visual novels |
| real-time | user actions immediately affect the generation stream |
| Stated output, from Runway's X post quoted across two passes: **continuous | |
| 720p video at 24 fps with 48,000 Hz audio**, **not restricted to a fixed | |
| duration**. |
Stated recipe, from one pass: fine-tune a base video model to the WorldPrompt format, post-train for autoregressive generation, distil for real-time speed. The base model is named only as Runway's "foundational audio-visual generation model" — no model name, no parameter count, no hardware requirement anywhere in what was read.
Compared To
- GWM Worlds — the predecessor this model is stated to build on, adding generated audio and rich subject and scene control. No page here: this wiki holds no release document for it.
- Gemini 3.8 Live — the nearest thing this wiki holds on the generative side, and the contrast is provenance rather than capability. Live Avatar carries SynthID on all generated audio and video; GWM Worlds 2 has no stated watermarking of any kind while emitting arbitrary-length photoreal video and audio. See Content Provenance (AI output marking).
- Cosmos-H-Dreams and Qwen-Drive-1.0-4B — world models in this wiki's other reading, built to train or evaluate an embodied policy rather than to be watched and steered. Neither is a comparator on any shared instrument, because none exists. See World Models.
Sources
- Runway Research, Introducing GWM Worlds 2 — not read,
runway.comanswersEGRESS_BLOCKED(source) (Runway) - Runway on X, 2026-09-03 — the quoted spec line (X)
- Latent Space, Runway's WorldPrompt and the Engineering of Real-Time Worlds,
2026-09-25 — not read,
www.latent.spaceanswersEGRESS_BLOCKED(Latent Space) - (AlphaSignal) (Spatial Insider)