Meta Releases Open-Source Llama 4 with 1 Trillion Parameters

Meta has released Llama 4, its most ambitious open-source language model to date, featuring over 1 trillion parameters in its largest configuration. The model, available for free download under a permissive license, matches or exceeds the performance of GPT-4 Turbo across most benchmarks — a development that could fundamentally shift the economics of the AI industry.

Model Variants

Llama 4 comes in four sizes to suit different deployment needs:

• Llama 4 Scout (8B): Designed for edge devices and personal use
• Llama 4 Pro (70B): Balanced performance for enterprise applications
• Llama 4 Ultra (405B): High-performance model for demanding tasks
• Llama 4 Behemoth (1T): Research-grade model pushing the frontier

Performance Benchmarks

The Llama 4 Ultra (405B) model is the most interesting from a competitive standpoint. It achieves 89.2% on MMLU, 94.1% on HumanEval, and 76.8% on GPQA — numbers that put it within striking distance of the best proprietary models, at zero API cost.

"Our goal has always been to democratize AI," said Mark Zuckerberg in a video announcement. "Llama 4 proves that open source can compete with — and in many cases beat — the most expensive proprietary alternatives."

Industry Impact

The release has immediate implications for the AI startup ecosystem. Companies that have built businesses around providing access to proprietary models now face an existential question: why would customers pay for API access when a comparable model is available for free?

Within hours of the release, several AI startups announced they would switch their backends from proprietary APIs to Llama 4. "This changes everything," said the CEO of a YC-backed AI startup. "Our inference costs just dropped by 80%."

The Open Source Debate Continues

Not everyone is celebrating. Safety researchers have raised concerns about releasing a model this powerful without restrictions. Meta has included safety guardrails in the model's system prompt, but these can be easily circumvented by fine-tuning — a reality that has drawn criticism from AI safety advocates.

Despite these concerns, the broader AI community has largely welcomed the release. Within 24 hours, Llama 4 had been downloaded over 500,000 times from Hugging Face, making it the fastest-adopted open-source model in history.