Meta Llama 4 Scout & Maverick: Open-Source AI Explained

Meta shipped Llama 4 Scout and Llama 4 Maverick on April 5, 2025. The two open-weight models are the first public Llama 4 releases. Scout packs 109 billion total parameters and a 10 million token context window. Maverick packs 400 billion total parameters with 17 billion active parameters and 1 million token context. Both models are available through Hugging Face and Meta AI. The release is significant because it puts near-frontier performance in hands that closed APIs often gate. Meta published benchmark numbers that challenge GPT-4o and Gemini 2.0 Flash on several tasks. For context on paid tier shifts, see major AI model tier changes.

Meta AI announced the models on its own homepage rather than a single model card. The primary source is Meta AI, and the weights sit on Hugging Face. Llama 4 Scout and Maverick use a mixture of experts architecture. Only 17 billion parameters activate per token for either model. That design keeps inference cost lower than a dense model of the same total size. It also changes how developers plan hardware. You still need serious GPU memory for either model, but active parameter counts matter for speed. The release date, April 5, 2025, came after months of smaller Llama 3 updates.

Why does this matter for open tool users? Maverick is Meta’s most capable open-weight general model at release. Meta reports Maverick scored 80.5 on MMLU Pro, slightly ahead of OpenAI GPT-4o at 80.4 and above Gemini 2.0 Flash. That is vendor-reported data, so independent tests may differ. Still, the gap between open weights and closed APIs has narrowed. Scout offers a 10 million token context window, which is rare at any price. This enables long document analysis, codebase review, and research tasks that would crush smaller context models. Licensing allows commercial use with limits, so businesses can self-host without a per-token API bill.

The launch lands during a messy period for free AI access. Several providers have tightened free tiers and pushed flagship models behind paid plans. Open weights offer a different route for teams that can manage infrastructure. Llama 4 is not a free hosted API, but the model files do not expire. For more on the June 2026 free tier shifts, see AI updates today. The tradeoff is real. You pay for compute, not for tokens. That can be cheaper or far more expensive depending on scale. Meta chose open weights as a strategic counter to closed competitors.

How Do the Top Options Compare?

Vendor-reported benchmarks and context figures as of April 5, 2025. GPT-4o context is 128,000 tokens. Closed model parameter counts are not disclosed. Running full context requires large GPU memory.

1. Llama 4 Scout , Long context document analysis and open research

Llama 4 Scout is the smaller of the two new Meta releases, but its context window is the headline. Scout carries 109 billion total parameters with 17 billion active parameters per token. That mixture of experts setup means the model does not fire all parameters for every input. The result is lower compute per query than a dense 109 billion parameter model. Scout supports a 10 million token context window. That is enough to ingest hundreds of pages of text, a full code repository, or a long research dataset in one prompt. No closed flagship model from OpenAI or Google offered 10 million token context at release. For open model options that avoid API costs, see best free AI models 2026. Meta positions Scout for retrieval, summarization, and agentic workflows that need very long memory. The model outperforms earlier open models of similar active size on Meta’s reported benchmarks. Meta claims Scout beats Gemma 3, Gemini 2.0 Flash, and Mistral 3.1 on several reasoning tasks. Independent verification is still light, so treat vendor benchmarks as a starting point. The downside is memory. A 10 million token context requires huge KV cache. Running Scout with full context may demand hundreds of gigabytes of VRAM. Quantization helps, but long context inference remains expensive. For most small teams, Scout is a cloud rental or a research project, not a laptop model. Scout uses the Llama 4 Community License. The license allows commercial use and fine-tuning, but it imposes restrictions for platforms with over 700 million monthly active users and for certain prohibited uses. This is open weights, not fully open source in the OSI sense. The distinction matters if your company wants to modify and redistribute without limit. Still, Scout gives long context researchers a direct alternative to closed APIs. You keep the weights and can run them on your own infrastructure.

Key strengths:

• ✅ 10 million token context window leads open and closed models at release
• ✅ 17 billion active parameters keep per-token compute lower than dense models
• ✅ Commercial use and fine-tuning allowed under community license
• ✅ Runs on your own hardware with no per-token API fees
• ❌ 109 billion total parameters still require serious GPU memory
• ❌ License not OSI open source due large platform restrictions
• ❌ Long context KV cache costs can be high

Who it’s for: Developers and researchers who need very long context analysis without closed API limits.

2. Llama 4 Maverick , General assistance, coding, and open-weight deployment

Llama 4 Maverick is Meta’s flagship open-weight general model in the April 5 release. It has 400 billion total parameters and 17 billion active parameters per token. The mixture of experts design spreads capacity across many expert modules. Maverick supports a 1 million token context window, below Scout but far above most closed chat models. Meta reports Maverick scored 80.5 on MMLU Pro, edging OpenAI GPT-4o at 80.4 and staying ahead of Gemini 2.0 Flash. It also posted competitive results on reasoning, coding, and math benchmarks. Those are vendor-reported numbers, so independent evaluations may differ. The practical point is that open weights now sit close to closed frontier models on common tests. Maverick is designed as a workhorse for general chat, coding, and multimodal tasks. Meta states the model is trained to handle text and image inputs, though the public release focus has been text. For developers watching coding tool costs, AI coding tools pricing GitHub Copilot usage based billing shows why self-hosted models matter. Maverick cannot match the polished ecosystem of a hosted Copilot plan, but it can run inside your own stack. You avoid per-seat and usage fees if you have the compute. The tradeoff is that 400 billion total parameters demand a lot of memory. Even with 17 billion active, you need enough VRAM to store the full model plus context. Expect multi-GPU servers or cloud instances. Benchmarks aside, Maverick matters because Meta released it under a commercial license. Businesses can fine-tune and deploy without paying a model API tax. The Llama 4 Community License has restrictions for very large platforms and prohibited uses, but it covers most independent developers and mid-size companies. The model is open weight, not fully open source. That honest limitation matters for research transparency. Still, Maverick gives teams a real option to own their model instead of renting it. The main downside is operational burden. Running a 400 billion parameter model in production is not a weekend project.

Key strengths:

• ✅ 400 billion total parameters with 17 billion active gives strong capability per token
• ✅ Reported MMLU Pro score of 80.5 edges GPT-4o
• ✅ Commercial license allows self-hosting and fine-tuning
• ✅ 1 million token context supports long document and code tasks
• ❌ 400 billion total parameters need multi-GPU hardware
• ❌ Vendor-reported benchmarks lack broad independent confirmation
• ❌ Open weights not fully open source under OSI definition

Who it’s for: Teams that want near-frontier open-weight performance and are willing to manage serious GPU infrastructure.

3. Llama 4 Weights on Hugging Face , Self-hosted deployment and commercial fine-tuning

The fastest way to start is not a single repo link. Go to Hugging Face or Meta AI and follow the official Llama 4 model pages. You must accept the Llama 4 Community License before downloading. That license is not the same as Apache 2.0 or MIT. It allows commercial use and fine-tuning, but it restricts platforms with over 700 million monthly active users unless Meta grants additional rights. It also includes acceptable use rules. The model weights are large. Scout takes roughly 200 gigabytes in full precision, and Maverick is far larger. Quantized versions can reduce that, but you still need high-VRAM GPUs. Self-hosting changes the cost equation. You pay for servers, not for tokens. For hobby projects, cloud GPU rental may cost more than a closed API. For production workloads, owning the model can lower unit costs at scale, but only if your team can handle deployment, monitoring, and inference optimization. The free tier landscape has shifted for hosted models. See AI API free tiers limits 2026 for what closed providers now offer. Open weights do not expire, so you are not tied to a vendor’s rate limit changes. That is the main appeal. Most developers should start with a quantized version on a rented multi-GPU instance before buying hardware. You will need tooling like vLLM, Hugging Face Transformers, or llama.cpp support. At the time of release, support in popular inference engines was still maturing. The honest downside is setup time. A closed API call takes seconds. Running a 109 billion or 400 billion parameter model takes real engineering. If you value control over convenience, the open-weight path is compelling. If you want fast prototyping, a hosted model may still win.

Key strengths:

• ✅ Direct download from Hugging Face or Meta AI without repo guesswork
• ✅ Commercial use allowed under Llama 4 Community License
• ✅ Full control over data and inference stack
• ✅ No per-token fees after infrastructure cost
• ❌ Large weights require high-VRAM GPUs and engineering time
• ❌ License limits very large platforms and some uses
• ❌ Inference tooling support may lag closed APIs

Who it’s for: Self-hosting teams that want model ownership and can manage GPU infrastructure.

4. Closed Hosted Rivals: GPT-4o and Gemini 2.0 Flash , Teams that want API convenience without GPU infrastructure

Closed models remain the default for most developers. OpenAI GPT-4o and Google AI Gemini 2.0 Flash offer hosted APIs with no local hardware. You send a request and pay per token. The downside is that access can change with pricing updates and free tier limits. Llama 4 Maverick reportedly edges GPT-4o on MMLU Pro at 80.5 versus 80.4, but that is one vendor-reported score. Closed models often win on ecosystem, tooling, and multimodal product polish. They also release updates without requiring you to download weights. Closed hosted models are simpler for small teams. You do not manage GPUs, inference servers, or model quantization. The cost model is different. You pay for every request, so high-volume use can become expensive. The June 2026 free tier shifts show how fast terms change. See free AI pricing changes June 2026 for details. Open weights like Llama 4 trade convenience for control. If your workload is low volume, a closed API is often cheaper than renting a GPU server. If your workload scales, self-hosting may pay off. The honest advice is to benchmark both paths. Meta’s release does not erase closed models. It gives you leverage and a fallback. Closed rivals still hold advantages in latency, managed safety, and multimodal features. Llama 4 Scout’s 10 million token context is ahead of GPT-4o’s 128,000 token window, so long context is a differentiator. For general chat and coding, Maverick is competitive but not clearly superior in independent tests. The right choice depends on your data, budget, and infrastructure.

Key strengths:

• ✅ Managed APIs require no GPU setup or maintenance
• ✅ Closed models have mature tooling and multimodal polish
• ✅ Pay per token can be cost effective at low volumes
• ❌ Pricing and free tier limits can change quickly
• ❌ No access to model weights for fine-tuning
• ❌ Context windows are shorter than Llama 4 Scout 10M token

Who it’s for: Developers who want fast integration and do not want to manage model infrastructure.

Frequently Asked Questions

What Should You Remember?

• Release date: Meta shipped Llama 4 Scout and Maverick on April 5, 2025.
• Parameters: Scout has 109 billion total and 17 billion active parameters. Maverick has 400 billion total and 17 billion active.
• Context: Scout offers 10 million tokens, Maverick offers 1 million tokens.
• License: Llama 4 Community License allows commercial use but restricts very large platforms.
• Access: Weights are available on Hugging Face and Meta AI after accepting the license.
• Benchmarks: Meta reports Maverick at 80.5 MMLU Pro, just ahead of GPT-4o at 80.4.
• Cost: Free weights, but self-hosting requires serious GPU hardware and engineering.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.