Llama 4

What It Is

Llama 4 is Meta's latest open-weight model family, released April 2025. Scout (109B total, 17B active, 16 experts) handles up to 10M tokens of context. Maverick (400B total, 17B active, 128 experts) is fine-tuned to 1M tokens. Both are natively multimodal with image understanding built into the architecture. The flagship Behemoth (2T total, 288B active) was announced but remains unreleased as of March 2026.

Why It Matters

Llama 4 is Emerging because the models are genuinely useful but the launch was badly damaged by a confirmed benchmark manipulation scandal. Departing Meta AI chief Yann LeCun confirmed in January 2026 that results were "fudged" — Meta used a non-public experimental variant for leaderboard submissions. CEO Zuckerberg reportedly "lost confidence in everyone involved," and the GenAI organisation was sidelined.

Despite the controversy, Scout and Maverick are available with downloadable weights and competitive pricing ($0.11/M input for Scout). For teams that need long-context multimodal capabilities with open weights, Llama 4 remains a practical option — you just can't trust the published benchmarks.

Key Developments

• Jan 2026: Departing Meta AI chief confirms benchmark results were "fudged" using a non-public model variant.
• Post-launch: GenAI organisation sidelined. Behemoth postponed indefinitely (11+ months past family launch).
• Independent testing: Real-world performance showed regression versus Llama 3 on coding tasks.
• Ongoing: Scout and Maverick weights available on Hugging Face for download and self-hosting.

What to Watch

Behemoth is the question mark. If Meta releases it with genuine benchmarks under new leadership, Llama 4's position improves significantly. If Behemoth remains vapour, the credibility damage compounds and teams will look to Llama 3 (proven) or DeepSeek R1 (competitive, MIT licensed) instead. The organisational restructuring at Meta AI is the signal to track.