Rumours

LLaMA 4 is “already done and fully AGI-capable,” according to leakers

Reviewing the rumour that Llama 4 has already been completed internally and is capable of AGI-level performance.

AI Rumour Watch

28 Nov 2025 — 2 min read

Claim: Llama 4 is complete and has achieved AGI-level performance.

Verdict: False — no AGI-capable model exists, and the rumour lacks credible verification.

Every six weeks the internet decides Meta has secretly built AGI and is just waiting for the right moment to unleash it. This week’s rumour claims that “Llama 4 is finished, fully AGI-capable, and outperforming every model on Earth,” based on—of course—a cropped screenshot, a Telegram channel, and someone who says they “know a guy at Meta.”

Let’s pull this one apart.

The Claim

The rumour asserts that:

Meta has completed Llama 4
The model “crosses into AGI territory”
It “easily surpasses GPT-4 and Claude 3”
Internal safety teams are “in panic mode”
Meta may “go public at any moment”

The favourite detail is that the model supposedly “solves unsolved maths problems effortlessly,” a phrase that usually translates to: “Someone saw a model do arithmetic.”

The Source

The earliest version of the rumour traces back to a Telegram channel known for posting a mixture of:

• Genuine model rumours
• Fabricated benchmark charts
• Crypto shilling
• And a lot of unearned confidence

The circulating screenshot:

• Has no metadata
• Uses a font that Meta does not use internally
• Repeats benchmark metrics from a 2023 Llama 2 leak
• Includes a category called “General Awareness”
• And has a watermark from a channel called “AITruthDrop”

A second version appeared on X, where it picked up momentum thanks to an account whose previous predictions include “GPT-5 is already conscious” and “Elon is building a god model in a warehouse.”

Our Assessment

Here’s what we can confirm:

• Meta has not announced or teased Llama 4 training completion.
• Meta’s public infrastructure capacity does not support the scale implied in the rumour.
• Meta’s actual research papers emphasise alignment, efficiency, and open release strategies—not AGI claims.
• No credible Meta insider has mentioned anything resembling “AGI capability.”
• The supposed benchmark chart uses axes and labels that do not exist in any ML research framework.
• The chart template is identical to one used in a 2024 YouTube thumbnail about “Secret AGI Projects.”
• The “AGI” claim is undefined, unsupported, and inconsistent with every public dataset or scaling law Meta has used to date.

So this “leak” is not a leak. It’s a hallucination with branding.

The Signal

Meta is working on:

• New Llama 3 and Llama 4 research directions
• Larger multimodal models
• More efficient training pipelines
• Reinforcement learning improvements
• Tools to improve reasoning and factual accuracy
• Open release strategies with safety review periods

But none of this points to:

• A completed Llama 4
• AGI
• A dramatic surprise announcement
• A sudden capability jump beyond known research
• Internal panic
• Or anything that matches the circulating images

The real signal is that Meta is iterating steadily, but with transparency, published research, and predictable milestones—not secret AGI breakthroughs leaking through Telegram.