Llama 4 is Meta's mixture-of-experts generation: Scout (109B total) and Maverick (400B total), both activating 17B parameters per token, both natively multimodal on text and images. Maverick still posts the highest MMLU among open models in mid-2026 roundups (85.5%), and Scout's headline 10M-token context claim remains the largest advertised window in open weights, with ~1M tokens the practically-validated range.
Llama's superpower is its ecosystem: every inference framework, fine-tuning library, and cloud platform supports it on day one, and the Llama 4 Community License — while not OSI-open (700M-MAU clause, branding requirements) — is free for virtually every company that isn't a hyperscaler. Amid the 2026 wave of Chinese frontier releases, Llama 4 remains the flagship American open-weight family and the safe institutional default.
Llama 4 specifications
| Vendor | Meta (United States) |
|---|---|
| Architecture | Mixture-of-Experts, natively multimodal (text + image) |
| Total parameters | Scout 109B / Maverick 400B |
| Active parameters | 17B (both) |
| Context window | Maverick 1M tokens; Scout up to 10M (claimed) |
| Modalities | text, image |
| License | Llama 4 Community License — free for most commercial use; requires a separate license above 700M MAU and carries branding/acceptable-use conditions |
| Release date | 2025-04-05 |
| Weights | Hugging Face · GitHub |
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| MMLU (Maverick) | 85.5% | highest MMLU among open models per mid-2026 roundups |
| MMLU Pro (Maverick) | 80.5 |
Running Llama 4 locally
Scout (109B/17B active) fits a single H100 at 4-bit (~60GB) — the most deployable long-context model of its class. Maverick (400B/17B) wants a multi-GPU node (~220GB at 4-bit).
What Llama 4 is best for
- The broadest tooling and ecosystem support in open weights
- Single-GPU long-context deployment (Scout on one H100)
- Organizations that prefer a US-based vendor for governance reasons
Frequently asked questions
Is Llama 4 really open source?
Llama 4 is open-weight under Meta's Community License: free for most commercial use, but with a 700M-monthly-active-user threshold requiring a separate Meta license, plus branding and acceptable-use conditions. It fails OSI's open-source definition — the practical impact for most companies is zero, but enterprises should review the terms.
What can Llama 4 Scout's 10M context actually do?
The 10M-token window is Meta's advertised architectural limit; independent evaluations validate strong recall to roughly 1M tokens, with degradation beyond. Even at 1M it is among the longest usable contexts in open weights, on a model that fits one H100.