Meta Mango: Everything We Know About Meta's Image & Video AI Model (2026)

September 19, 2026 · 6 min read

Meta Mango is the reported codename for an image and video generation model in development at Meta Superintelligence Labs — not yet public, not officially announced, and first reported by the Wall Street Journal in December 2025. A sibling text model is codenamed Avocado. If you're searching for a download link or a demo, stop: as of September 2026 there isn't one. Here's what's genuinely known, what's inference, and why this model matters more than most unreleased models do. One note on sourcing: everything below traces to reporting about Meta's internal plans, so treat dates and capabilities as provisional until Meta says otherwise.

What is actually known

• It exists as a project. The WSJ reported in December 2025 that Meta is developing an image and video model internally codenamed "Mango," alongside a text-based model codenamed "Avocado."
• Who's building it. Meta Superintelligence Labs — the unit Meta formed around Scale AI co-founder Alexandr Wang, with Chris Cox involved at the product level. Meta restructured its AI org and spent heavily hiring for this effort through 2025.
• The original target slipped. The reported plan was a release in the first half of 2026. It's September, and no public model has shipped — which suggests either a delay, a quiet soft launch, or a scope change. Meta has not commented on the codenames.

Everything else you've read — resolutions, benchmarks, pricing — is speculation. There are no public specs.

Why Mango matters more than another unreleased model

Models are common; distribution is not. Meta owns Instagram, Facebook, WhatsApp and Threads — the places where AI images and video actually get consumed at scale. A competitive in-house image/video model doesn't need to win benchmarks to change the market; it needs to be good enough and free inside apps with billions of users. That's the scenario that should worry the standalone generators: not "Meta's model beats Midjourney," but "Meta's model is already in the composer where the post gets written."

There's also a strategic tell in the pairing. Mango (visual) plus Avocado (text) is a multimodal foundation strategy, not a feature — the same shape as Google's Gemini/Veo stack and OpenAI's GPT/Sora stack. Meta is the last of the big three without a competitive generative media model, and its Llama image efforts lagged badly enough that the superintelligence lab was the correction.

What it means for creators and marketers

• Expect free or bundled access. Meta's pattern with AI features is to give them away inside its apps to drive engagement. If Mango ships in-feed, casual generation gets commoditized further — the value moves to consistency, control and commercial rights, which is where studios earn their keep.
• Expect a training-data fight. Meta's history of training on user content means Mango will arrive with opt-out debates attached. If you generate commercially, watch the licence terms, not the demos.
• Don't wait for it. A slipped H1 target with no public demo means Mango is a 2027 story at the earliest for practical purposes. The frontier it has to beat — current image and video models — is available today.

What you can use right now

While Mango stays internal, the current generation of models is not standing still. On the image side, GPT Image 2.5, Nano Banana 2 and Seedream 5 Pro all run in Hermosso's Create studio from one prompt box. On the video side, Seedance 2.5 now generates 30-second clips natively — we wrote up how to make 30-second AI videos — and Kling 3 does native 4K. New accounts get 750 free credits, which is enough to benchmark the real frontier instead of reading rumours about the next one. We'll update this article the moment Meta makes Mango official.

Made with Hermosso