Llama 4 Scout
Last updated: August 10, 2026
Specs
| Vendor | Meta |
|---|
| Released | April 5, 2025 |
|---|
| Context window | 10M tokens |
|---|
| Input | $0.1 / M tokens |
|---|
| Output | $0.3 / M tokens |
|---|
Benchmarks
| mmlu | 80.4 |
|---|
| gpqa | 49.1 |
|---|
| humanEval | 76.3 |
|---|
| arenaElo | 1278 |
|---|
| sweBench | 35 |
|---|
| sweBenchPro | 25 |
|---|
| liveCodeBench | 50 |
|---|
| aiderPolyglot | 45 |
|---|
| terminalBench | 20 |
|---|
| mmluPro | 65 |
|---|
| hle | 6 |
|---|
| aime | 50 |
|---|
| math500 | 78 |
|---|
| mmmu | 60 |
|---|
Strengths
- 10M-token context window, unbeaten for bulk docs
- Runs on a single H100 with heavy quantization
- Open weights for self-host
Weaknesses
- Lower benchmark scores than Maverick
- Vision weaker than closed frontier models
- Quality drops significantly on deep reasoning
Best for
- Needle-in-the-haystack retrieval on huge corpora
- Self-hosted long-doc analysis
- Cost-sensitive summarization
Compare with other AI models