Skip to content
Bud's Wiki

Gemini 'Omni' leak: new omni model spotted on the video-generation tab — TestingCatalog

TL;DR

Section titled “TL;DR”

@testingcatalog (AI News) reports a leaked UI string from Gemini’s video-generation tab — “Start with an idea or try a template. Powered by Omni.” — suggesting a new Google omni model is being A/B tested as the engine behind Gemini’s video generator, displacing the current Veo-based pipeline (internal name “Toucan”). The post frames Omni as potentially “the first top-tier Omni model with video output” if it ships, and positions it as a Google I/O 2026 candidate. Treat as rumor: the only artifact is a UI screenshot, no model card, weights, or benchmarks.

Key claims

Section titled “Key claims”
  • A new tab string in Gemini’s video-generation UI reads “Powered by Omni,” interpreted by the reporter as the codename of a new model replacing Veo on that surface [tweet body].
  • “Toucan” is identified as the internal name of the current video generation tool powered by Veo, against which Omni is positioned [tweet body].
  • The reporter speculates Omni would outperform Veo 3.1 if shipped, and would be “the first top-tier Omni model with video output” — i.e. a unified model whose output modalities include video, not just text+image [tweet body].
  • Targeted reveal venue is Google I/O 2026 [tweet body].

Method

Section titled “Method”

A UI-string leak from the Gemini web app. The only primary artifact is a screenshot of the video-generation tab. No technical detail, no benchmark, no claim about architecture. Status: speculative.

Results

Section titled “Results”

None reported. 158K views, 30 replies, 92 retweets, 980 likes at the time of capture — i.e. broadly seen rumor, not validated.

Why it’s interesting

Section titled “Why it’s interesting”

If accurate, “Omni” as a single Google model spanning text + image + audio + video output would be the first frontier-scale entrant in the cohort tracked under Unified Multimodal Models where the output modalities include video — currently OmniWeaving (OmniWeaving: Towards Unified Video Generation with Free-form Composition and Reasoning) is the only filed unified-output model with video on the page, and it’s a Decoupled-then-E2E stack rather than a from-scratch omni. It would also be the closed-but-API-accessible counterpart to recent open omni releases (Qwen3.5-Omni, LongCat-Next, Ming-flash-omni, LTX-2) that Open foundation-model releases benchmarks against. Worth re-checking around Google I/O 2026; until then, this is a rumor, not a baseline.

See also

Section titled “See also”