Meta has released Llama 4, its first natively multimodal Mixture-of-Experts AI model family, with two downloadable open-weight versions that let developers customize and deploy AI without paying API fees. The announcement matters because it signals Meta's commitment to open-source AI at a moment when corporate America is rapidly shifting away from expensive closed models like those from OpenAI and Anthropic.

What Makes Llama 4 Different From Previous Versions?

Llama 4 introduces two major architectural changes that separate it from Llama 3. First, it moves to a Mixture-of-Experts design, meaning only selected parts of the model activate for each task rather than using all parameters every time. This allows Meta to build much larger models while keeping active compute lower. Second, Llama 4 is natively multimodal, handling text and image tokens inside a unified model backbone rather than bolting vision capabilities onto a text-only system.

The practical benefit for users is more natural image reasoning, document analysis, chart interpretation, and visual question answering. For developers, the appeal is clear: they can download these models, customize them for specific tasks, and avoid paying per-token fees to API providers.

Scout Versus Maverick: Which Model Should You Choose?

Meta released two downloadable open-weight models with very different strengths. Scout is the smaller, more deployable option with 17 billion active parameters and support for up to 10 million tokens of context, roughly equivalent to processing 100,000 words at once. Maverick is the more ambitious general-purpose assistant, also with 17 billion active parameters but 128 routed experts and around 400 billion total parameters, making it significantly heavier to serve.

Behind both models sits Llama 4 Behemoth, a massive 288-billion-parameter teacher model that Meta used to distill quality into the smaller versions. Behemoth is not a normal downloadable model; it exists primarily for research and explains why Scout and Maverick feel stronger than their active parameter counts suggest.

How to Evaluate Llama 4 for Your Organization

• Local Deployment Needs: Scout becomes plausible with heavy quantization on a single NVIDIA H100 GPU, but consumer GPUs with 8GB to 16GB of memory are too small. If you need private local inference on a workstation, Llama 4 may be a future target rather than an immediate option.
• General Assistant Tasks: Maverick is positioned for image understanding, creative writing, multilingual use, coding, and reasoning, but it requires hosted infrastructure or multi-GPU servers for realistic performance. Most teams evaluating Llama 4 should start with cloud-based API access rather than local deployment.
• Long-Context Workflows: Scout's 10-million-token context window is genuinely useful for multi-document work, large codebase reading, and long-context retrieval tasks, though quality still depends on retrieval behavior, attention implementation, and prompt structure.

The model cards published by Meta include concrete parameter and context details, making it easier for developers to test the exact model and provider they plan to use rather than relying on launch benchmarks alone.

Why Corporate America Is Abandoning Expensive AI APIs

The shift toward open-weight models like Llama 4 reflects a broader cost crisis in enterprise AI. AT&T, one of the largest US telecommunications companies, turned to open models after AI costs from providers like OpenAI and Anthropic skyrocketed. By May 2026, open models accounted for 20 percent of AT&T's AI use. That figure has since risen to 40 percent and may jump to 60 percent in coming months, with the company saving up to 80 percent on AI costs compared with earlier in the year.

"We believe it could go much, much higher," said Andy Markus, AT&T's chief data and AI officer, discussing the company's shift toward open models.

Andy Markus, Chief Data and AI Officer at AT&T

AT&T is not alone. Companies such as Airbnb and Deloitte are also turning to open AI models that are easy to customize and cheaper to use. According to user data from OpenRouter, a platform that lets people choose different AI models to complete tasks, open models accounted for 58 percent of AI use last month, up from just 10 percent a year ago.

The Competitive Threat to OpenAI and Anthropic

The rise of open models poses a significant challenge to OpenAI and Anthropic, both of which are heading toward blockbuster initial public offerings. These companies have to charge premium prices because they spend billions on research and development and computing power. Meanwhile, some Chinese open models are now 80 to 90 percent as powerful as those from OpenAI and Anthropic while costing as little as 20 percent of the price.

"Mazdas do a fine job getting you where you need to go, and Maseratis are simply not worth the investment," said Jerry Tang, CEO of Atlas Cloud, comparing expensive closed models to open-source alternatives.

Jerry Tang, CEO of Atlas Cloud

The popularity of open models reflects a broader debate in Silicon Valley and Washington about control over AI. Dario Amodei, CEO of Anthropic, has argued that AI should be tightly regulated to prevent national security concerns. Other tech leaders, including NVIDIA CEO Jensen Huang and Meta CEO Mark Zuckerberg, have said open-source technology is crucial for a democratic, balanced AI ecosystem.

Most US companies still use a combination of open and closed models. Closed models remain the best for heavy-duty tasks like coding and image and video generation, while open models often excel at simple, specialized tasks. However, AT&T is now taking open-weights models and customizing them to make new tools that can transcribe calls and facilitate customer service.

What Are the Practical Limitations of Llama 4?

Despite its strengths, Llama 4 is not a normal laptop model. Meta says Scout can fit on a single NVIDIA H100 GPU with Int4 quantization, while Maverick can run on a single H100 DGX host or distributed inference for better efficiency. That is very different from saying it runs comfortably on an 8GB or 12GB gaming card.

Additionally, Maverick's leaderboard performance came under scrutiny when TechCrunch reported that the LM Arena result used an experimental chat-optimized variant, while the widely available downloadable model ranked lower once evaluated separately. This does not make Llama 4 useless, but it does mean developers should test the exact model and provider they plan to use rather than relying on one launch benchmark.

For developers, the official starting point is llama.com downloads, with the two main Hugging Face repositories being meta-llama/Llama-4-Scout-17B-16E-Instruct and meta-llama/Llama-4-Maverick-17B-128E-Instruct. Commercial users should read the Llama 4 Community License and acceptable use policy before shipping a product, as Llama is open-weight but the license is not the same as a no-conditions Apache 2.0 release.

For ordinary users, the quickest entry point is Meta AI, which also integrates Llama 4 into WhatsApp, Messenger, and Instagram Direct experiences. Availability can vary by country, account, and product rollout, so the web version and app integrations may not expose every model option in the same way.