Fact Check: There Is No Llama 5

While compiling this round of model news we ran into a widely-syndicated claim that Meta released Llama 5 on 8 April 2026 with 600 billion parameters, a 5-million-token context window, and a novel "Recursive Self-Improvement" capability letting the model correct its own reasoning mid-task without external feedback.

None of that appears to be true. We could not find it in Meta's own record, and three things contradict it:

1. Meta's AI blog documents no Llama 5 release. The most recent Llama generation on it is Llama 4.
2. In April 2026 Meta Superintelligence Labs released Muse Spark - and notably as a proprietary model, a departure from the open-weight approach Llama was known for. Coverage at the time framed this as Muse Spark replacing the Llama line, not accompanying a Llama 5.
3. Prediction markets asking "when will Meta release Llama 5?" were still open and unresolved at the time of writing, which would be impossible had it shipped four months earlier.

The "recursive self-improvement" detail is the tell. It is the kind of specific-sounding but unfalsifiable capability claim that machine-generated content farms produce, and it does not correspond to any documented Meta research.

What is actually true

Llama 4 was released on 5 April 2025 - Scout and Maverick, natively multimodal with a mixture-of-experts architecture. Llama 4 Behemoth was announced but never released; it was still in training. Llama 4 Scout, with its 10M-token context window, remains on our roster and remains the latest generally-available open-weight Llama.

Earlier reporting had Meta planning an interim Llama 4.X or 4.5 before the end of 2025, and a flagship codenamed Avocado targeted at Q1 2026. Neither shipped under those names.

We are publishing this as its own item because the false version is detailed, confident and well-indexed - exactly the profile that gets repeated. Llama 4 Scout is Meta's current open-weight Llama, and Meta's newest model is not open-weight.