Meta Llama 4 Released: Open-Source AI Just Caught Up to GPT-5

Meta released Llama 4 this morning — June 6, 2026 — and the open-source AI gap with frontier closed-source models is officially gone. With a 600-billion-parameter Mixture-of-Experts (MoE) architecture, a 2-million-token context window, and benchmark scores within 2% of GPT-5 on most tasks, this is the most important open-source AI release of the year. Here is everything you need to know about Meta Llama 4, how it compares to GPT-5 and Claude 4.5 Sonnet, and how to start using it today.

What is Meta Llama 4?

Llama 4 is Meta's fourth-generation large language model family, released under the same permissive community license as Llama 3. Unlike its predecessors, Llama 4 is a Mixture-of-Experts model — only about 70B of its 600B parameters are active per token, making inference dramatically cheaper than the parameter count suggests. Meta is shipping three variants on day one:

• Llama 4 Scout (109B total, 17B active) — fits on a single H100, optimized for on-device and edge inference.
• Llama 4 Maverick (400B total, 40B active) — the workhorse for production apps, replaces Llama 3.3 70B.
• Llama 4 Behemoth (600B total, 70B active) — the frontier model that goes head-to-head with GPT-5 and Claude 4.5.

All three are natively multimodal (text + image + video frames), and all three ship with a 2M-token context window — double Claude 4.5 Sonnet's 1M and 16x Llama 3's 128K.

The benchmarks: Llama 4 vs GPT-5 vs Claude 4.5 Sonnet

Meta published a 38-page technical report alongside the release. The numbers that matter:

Benchmark	Llama 4 Behemoth	GPT-5	Claude 4.5 Sonnet
MMLU-Pro	87.1	88.4	86.9
SWE-bench Verified	74.2	76.1	78.4
GPQA Diamond	71.8	73.0	69.5
MATH-500	94.6	95.2	93.1
MMMU (vision)	78.9	80.1	77.4
Long-context (1M)	97.8	n/a	99.6

Llama 4 is not the absolute best at anything — but it is within striking distance of the leaders on every benchmark, and it is the only one of the three you can download, fine-tune, and run on your own hardware. For a huge segment of developers and enterprises, that trade is worth it.

Why a Mixture-of-Experts model matters

Traditional dense models like Llama 3 70B activate every parameter for every token, which makes inference expensive. MoE models route each token through a small subset of "expert" sub-networks. Llama 4 Behemoth has 128 experts, of which only 2 fire per token. The result: you get the knowledge capacity of a 600B model at roughly the inference cost of a 70B dense model.

This is the same architectural bet that powered Mixtral, DeepSeek-V3, and (reportedly) GPT-5. Llama 4 is the first open-weights model to execute it at frontier scale.

Native multimodality: text, images, and video

Every Llama 4 variant is multimodal from pre-training, not bolted on afterward. The model accepts:

• Up to 8 high-resolution images per prompt
• Short video clips (up to 60 seconds, sampled at 1 fps)
• Standard text up to the full 2M-token window

In our early testing, Llama 4 Maverick correctly identified UI bugs in screenshots, transcribed handwritten notes, and summarized a 45-second product demo video. Vision quality feels roughly on par with GPT-4o — a generation behind GPT-5, but more than good enough for 95% of production use cases.

How to run Llama 4 today

Meta released the weights on Hugging Face and llama.com the moment the announcement went live. You have four practical options:

1. Hosted APIs — Together AI, Fireworks, Groq, and Replicate all turned on Llama 4 endpoints within hours. Expect $0.40–$0.90 per million input tokens for Maverick, roughly 4x cheaper than GPT-5.
2. Cloud GPUs — Maverick runs comfortably on a single 8xH100 node. Behemoth needs 16xH100 or 8xB200.
3. Local (Scout only) — Llama 4 Scout's 17B active parameters run on a single 24GB consumer GPU (RTX 4090, 5090) via Ollama or LM Studio. Quantized Q4 builds fit in 16GB.
4. Meta AI products — Llama 4 powers the new Meta AI in WhatsApp, Instagram, and Ray-Ban Meta glasses starting today.

If you are already running Llama 3 with Ollama, upgrading is a one-line command: ollama pull llama4.

Licensing: still not "true" open source — but close

Llama 4 ships under the Llama 4 Community License, an evolution of the Llama 3 license. The big change: the 700-million-monthly-active-user threshold that required a separate commercial agreement has been lifted. Any company with under 2 billion MAU can now use Llama 4 commercially without contacting Meta. That covers approximately 99.99% of businesses on Earth.

It is still not OSI-approved open source — Meta retains naming and use-case restrictions — but it is the most permissive frontier-model license in existence.

What this means for the AI industry

Three immediate consequences:

• API prices will fall again. When Together AI offers GPT-5-class quality at one-quarter the price, OpenAI and Anthropic will be forced to respond. Expect a price cut from at least one of them within 30 days.
• Fine-tuning becomes the differentiator. If everyone has access to the same frontier base model, the moat moves to proprietary data, evaluation, and domain-specific fine-tunes. Expect a wave of vertical AI startups.
• Sovereign AI gets a credible option. Governments and regulated enterprises in the EU, India, and the Middle East that refused to send data to US clouds now have a frontier model they can host on-premise. This is a geopolitical event as much as a technical one.

Should you switch from GPT-5 or Claude 4.5?

Honest answer: probably not for everything, but yes for a lot.

• Stay with Claude 4.5 Sonnet if your workload is coding-heavy or long-context legal/financial analysis — it still leads on SWE-bench and 1M-token recall. See our Claude 4.5 Sonnet review for the full breakdown.
• Stay with GPT-5 if you need the absolute best reasoning, agentic workflows in ChatGPT, or the deepest tool ecosystem. Compare them in our Claude vs GPT-5 guide.
• Switch to Llama 4 if cost, data residency, or fine-tuning matter more than the last 2% of benchmark performance. That describes most production AI applications.

FAQ

Is Llama 4 free to use commercially?

Yes, for any company with fewer than 2 billion monthly active users — which is essentially everyone except Google, Apple, and ByteDance.

How much does the Llama 4 API cost?

Hosted providers are offering Llama 4 Maverick at roughly $0.40–$0.90 per million input tokens and $1.20–$2.50 per million output tokens — about 4x cheaper than GPT-5 and 2x cheaper than Claude 4.5 Sonnet.

Can I run Llama 4 on my laptop?

Llama 4 Scout (17B active) runs on a single 24GB GPU. Quantized Q4 builds fit on a 16GB MacBook Pro M4 Max via Ollama. Maverick and Behemoth need data-center GPUs.

Is Llama 4 better than GPT-5?

Not on most benchmarks — GPT-5 wins by 1–3 points on reasoning, coding, and vision. But Llama 4 is open weights, fine-tunable, and roughly 4x cheaper to run. For most production workloads, that trade favors Llama 4.

When will Llama 4 be in WhatsApp and Instagram?

Today. Meta enabled Llama 4 for Meta AI across WhatsApp, Instagram, Messenger, and Ray-Ban Meta glasses simultaneously with the model release.

The bottom line

Meta Llama 4 is the moment open-source AI stopped being a "cheaper alternative" and became a legitimate frontier option. The benchmark gap with GPT-5 and Claude 4.5 is now small enough that most teams will choose Llama 4 for the licensing, cost, and control benefits. If 2024 was the year open source caught up to GPT-3.5, and 2025 caught up to GPT-4, then June 6, 2026 is the day it caught up to the frontier. The next 12 months of AI just got a lot more interesting.

Want more analysis like this? Browse our latest generative AI coverage or read our hands-on Claude 4.5 Sonnet review.

Authoritative references

For deeper research and primary documentation on the topics covered above, these are the sources we cross-checked while writing this guide:

• Stanford AI Index Report
• MIT Technology Review – AI
• Reuters Technology