Llama 4 Maverick — Benchmarks, Capabilities, and Use Cases
Llama 4 Maverick is a open-weight AI model from Meta (17B/400B MoE (128 experts)), classified as Established tier with a 1M tokens context window. Explore bench...
4 min read
Llama 4 Scout is a proven AI model still widely deployed in production systems, developed by Meta and released in April 2025. As an open-weight model, its trained weights are publicly available — you can self-host it, fine-tune it on proprietary data, or run it on-premise without API dependencies. Its 10M-token context window is one of the largest available — capable of processing book-length documents, entire code repositories, or very long conversation histories in a single prompt.
On the FlowHunt AI Leaderboard it is one of 24 tracked models, evaluated across 4 benchmarks. Standout scores: 50.3% on MATH-500 (#4 of 4); 32.8% on LiveCodeBench (#5 of 5); 52.2% on MMLU-Pro (#8 of 8).
Llama 4 Scout was released alongside Maverick on April 5, 2025 as the smaller, highly accessible variant of the Llama 4 generation — and immediately attracted attention for its context window: 10 million tokens, the largest of any model on this leaderboard. The 10M-token window was achieved through modifications to the attention mechanism and positional encoding that enabled the model to maintain coherence across documents far longer than any prior open-weight model could handle. Scout (17B active / 109B total, 16 experts) became the default model for retrieval-augmented generation pipelines that needed to operate over very large document collections without chunking — processing entire legal case histories, codebases, or scientific literature corpuses in a single prompt.
Note: 10M token architectural context. Single H100 (Int4). DocVQA 94.4%.
- Released
- April 2025
- Parameters
- 17B/109B MoE (16 experts)
- Context
- 10M tokens
- Weights
- Open
- Tier
- Established
- Provider
- Meta
Llama 4 Scout Benchmark Rankings
Rank among all 24 models tracked on this leaderboard. SR = self-reported · 3P / AA / C = independently verified.
| Benchmark | Score | Rank | Source | What it measures |
|---|---|---|---|---|
| MMLU-Pro | 52.2% | 8 of 8 | SR | Hard knowledge reasoning, 12K questions |
| LiveCodeBench | 32.8% | 5 of 5 | SR | Competitive programming, continuously updated |
| MATH-500 | 50.3% | 4 of 4 | SR | Hendrycks math (500 problems) |
| AA Intelligence Index | 14.0 pts | 9 of 9 | AA | Artificial Analysis composite score (independent) |
Detailed Benchmark Scores
MMLU-P
SR
52.2%
Hard knowledge reasoning, 12K questions
LiveCode
SR
32.8%
Competitive programming, continuously updated
MATH
SR
50.3%
Hendrycks math (500 problems)
AA Idx
AA
14.0 pts
Artificial Analysis composite score (independent)
Llama 4 Scout vs. Established Peers
Head-to-head benchmark comparison with other Established-tier models. Higher is better for all metrics.
| Benchmark | Llama 4 Scout | GPT-4o | Llama 4 Maverick |
|---|---|---|---|
| MMLU-P | 52.2% | 72.6% | 80.5% |
| LiveCode | 32.8% | — | 43.4% |
| MATH | 50.3% | — | — |
| AA Idx | 14.0 pts | — | — |
About Meta
Meta Platforms (formerly Facebook) is one of the world's largest technology companies, headquartered in Menlo Park, California, and best known for Facebook, Instagram, WhatsApp, and Threads. Through Meta AI (formerly FAIR), the company has become the most prolific publisher of open-weight large language models via the Llama family. The Llama 4 series — Maverick, Scout, and Behemoth — were released under permissive community licences and remain the most downloaded open-weight models on platforms like Hugging Face. Meta's AI strategy is distinct among major labs: rather than operating a commercial API, Meta releases models openly for research and commercial use, integrates generative AI features across its own applications, and has publicly advocated for open-weight releases as the most democratic approach to AI development.
Use Cases
What to Use Llama 4 Scout For
- Self-Hosted & Private Deployments
Llama 4 Scout is open-weight (17B/109B MoE (16 experts) parameters) — its trained weights are publicly available for download and self-deployment. Run it on your own GPU hardware or private cloud to ensure data never leaves your environment, eliminate per-token API costs at scale, or fine-tune the model on proprietary datasets.
- Long-Document & Codebase Analysis
With a 10M-token context window, Llama 4 Scout can ingest entire software repositories, long legal contracts, book-length research reports, or extensive conversation histories in a single prompt — enabling deep document Q&A, cross-file codebase reasoning, and comprehensive summarisation without chunking heuristics or RAG pipelines.
Strengths & Limitations
Strengths
- Self-hosted & data-private deployments (open-weight)
- Very long documents and large codebases (10M context)
Limitations
- Significantly behind 2026 frontier models on agentic coding and complex reasoning
Browse the Full AI Model Leaderboard
Compare all 24 models across 11 benchmarks — sortable, sourced, and updated June 2026.
Frequently asked questions
Learn more
Llama 3.3 70B — Benchmarks, Capabilities, and Use Cases
Llama 3.3 70B is a open-weight AI model from Meta (70B dense), classified as Established tier with a 128K tokens context window. It is available directly in Flo...
4 min read
Llama 4 Scout AI: Performance Analysis Across Multiple Tasks
An in-depth analysis of Meta's Llama 4 Scout AI model performance across five diverse tasks, revealing impressive capabilities in content generation, calculatio...
4 min read
AI
Llama 4
+8