Meta Just Released Llama 4 as Full Open Source. The AI Business Model Just Changed
Meta released Llama 4 under Apache 2.0 license with no commercial restrictions. At 405B parameters, it matches GPT-5.5 on most benchmarks. The proprietary AI model business just got a lot harder to justify.
Meta released Llama 4 yesterday. Full weights. Apache 2.0 license. No commercial restrictions. You can download it right now from Hugging Face or Meta's own repo.
The 405B-parameter model matches GPT-5.5 on MMLU, HumanEval, and most reasoning benchmarks. It's slightly behind Claude Opus 4.7 on coding and long-form writing. The 70B variant fits on a single high-end GPU with quantization.
Architecturally, Llama 4 uses a mixture-of-experts design with 16 experts and 2 active per token, so the 405B model behaves like a 50B model at inference time. Compared to Llama 3.1's dense architecture, inference runs about 3x faster and costs 4x less per token while scoring 8 to 12 points higher. Key improvements: a native 256K context window, multilingual support across 30 languages, and native tool-use for function calling and structured JSON output.
Against the proprietary frontier, Llama 4 scores 78.2 on MMLU-Pro versus GPT-5.5's 79.1 and Claude's 80.3, close enough that most applications will not notice. On HumanEval it scores 91.2% versus Claude's 94.8%. For code, structured output, and multilingual processing, Llama 4 is in the same tier as the proprietary leaders. For creative writing, Claude holds a meaningful lead.
The part that matters
The license. Apache 2.0. That means you can fine-tune it, build products on it, sell access to it, and never pay Meta a cent. No revenue sharing. No "you must display 'powered by Llama' on your app." No loophole where the license changes if you get too big.
This is meaningfully different from what other companies call "open source." Google's Gemma models use a custom license with usage restrictions. Mistral's top models are Apache 2.0, but the company also sells proprietary premium versions. Microsoft's Phi models are MIT-licensed but the company does not train anything near frontier scale. Meta is the only organization training frontier-scale models and releasing them under a standard, unrestricted open source license.
Zuckerberg's accompanying post was short. The key line: "Open source AI is catching up faster than anyone predicted. We think the ecosystem around Llama will create more value than selling API access ever would."
He might be right. Or he might be rationalizing the fact that Meta couldn't compete with OpenAI and Anthropic on model quality and chose a different game. Either way, the outcome is the same: 405 billion parameters of frontier-grade AI, free to use, for anyone.
The open source AI movement just got real power
The open source AI community has been fighting an uphill battle since ChatGPT launched. Llama 3 was good but behind GPT-4. Llama 3.1 closed some gap. Llama 4 genuinely changes the conversation: when you can download something matching GPT-5.5 on benchmarks, the question shifts from "can open source compete?" to "why pay for API access?"
This has immediate implications. Startups control their own inference stack, free from OpenAI's pricing changes. Regulated industries fine-tune on proprietary data without sending it to a third party. The research community gets a GPT-5.5-class model they can actually open up and study.
Who this hurts
OpenAI first. The $200/month Pro tier is built on GPT-5.5 access. When a free model matches your flagship, $200 becomes harder to justify. OpenAI still has advantages: the chat interface, GPT Store, DALL-E, voice mode, the brand. But the technical moat just got narrower.
Anthropic less directly. Claude's advantage is writing quality and careful reasoning, not raw benchmark scores. Nobody picks Claude because it is cheaper. They pick it because the writing is better, and Llama 4's writing is not better. Anthropic's moat might hold longer.
The biggest winner: developers. Llama 4 inference costs on Groq or together.ai run about 80% cheaper than GPT-5.5 API calls. At scale, that pays for entire engineering teams.
What this doesn't mean
It doesn't mean training frontier models is free. Llama 4 reportedly cost Meta around $400 million. Meta can afford that because it makes $160 billion a year from ads. The compute cost is not a line item they need to recoup.
It also doesn't mean open source automatically wins. OpenAI and Anthropic can still ship faster and build products open source cannot match. The GPT Store has thousands of specialized agents. Claude has managed agents with Dreaming, a feature requiring infrastructure beyond model weights.
Meta's strategy is worth examining honestly. Making Llama open source is not charity. Every developer building on Llama builds on Meta's stack. Every startup deploying Llama is one fewer locked into a proprietary API. Meta loses nothing when API prices drop. Its competitors do. The strategy is generous and ruthless, and it works because Meta's advertising business is orthogonal to the AI API business. Each release permanently lowers the ceiling on what anyone can charge for comparable capability.
But the pricing argument is gone. A year ago, frontier AI was arguably worth $200/month because the alternative was dramatically worse models or nothing at all. After Llama 4, the alternative is a model roughly as good for roughly free. That changes what premium AI companies can charge for.
Practical use cases
If you have GPU compute, Llama 4 shines in three areas. Document processing at scale: the 256K context window and multilingual support handle entire legal contracts or research papers. Internal knowledge bases: companies fine-tune on internal docs and deploy a private ChatGPT equivalent that never sends data outside. Code review automation: slightly behind Claude on coding benchmarks, but costs a fraction of API alternatives.
The 70B variant runs on a single consumer GPU with 4-bit quantization. That puts frontier-adjacent AI on hardware individuals can actually own. The psychological difference between "I use AI" and "I run AI" matters.
I have been following open source AI since the first Llama weights leaked in February 2023. Watching a model anyone can download now match the best closed-source systems is genuinely remarkable. If the trajectory holds, the subscription-based AI business built on raw model superiority has maybe 18 to 24 months of runway. OpenAI and Anthropic are not going anywhere, but their advantages are shifting from model quality to user experience and platform integration. The open source movement closed the model quality gap. Whether it closes the platform gap is the next question.
One thing I'm watching
Meta said Llama 4 was trained on "publicly available data" and that the training corpus would be published "in the coming weeks." If they actually release the dataset composition, it'll be the most transparency any major lab has provided about training data. If they don't, the "publicly available" language will get scrutinized hard.
The weights are up. The benchmarks are real. The license is real. Go grab it and see for yourself.