Llama 4 Maverick 17B 128E Instruct FP8 · Llama-4-Maverick-17B-128E-Instruct-FP8
Meta · 5 Apr 2025
benchmarks
Meta's Llama 4 Maverick FP8 is an instruction-tuned, multimodal mixture-of-experts model that accepts text and images and generates text. It has 17 billion active parameters across 128 experts, approximately 400 billion total parameters, and a documented one-million-token context window.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 69.8%Apr 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 43.4%Apr 2025 | View |
| MMLU-ProBroad expert knowledge and reasoning | 80.5%Apr 2025 | View |
| MMMUMultimodal (image + text) understanding | 73.4%Apr 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 20.6%8 Apr 2025 | View |
| FrontierMath | 0.7%8 Apr 2025 | View |
| MATH Level 5 | 73%8 Apr 2025 | View |
Weighted mean of 7 verified benchmark results: GPQA Diamond 69.8%, LiveCodeBench 43.4%, MMLU-Pro 80.5%, MMMU 73.4%, OTIS Mock AIME 2024-25 20.6%, FrontierMath 0.7%, MATH Level 5 73%.
Capabilities
- Processes text and image inputs together for visual question answering and image-grounded assistance.
- Supports a documented context window of one million tokens.
- Generates instruction-following conversational text in 12 supported languages.
- Supports coding assistance and text-based reasoning tasks.
- Provides downloadable FP8 weights for deployment on compatible inference infrastructure.
What improved
- Adds native image understanding compared with the text-only Llama 3.3 70B model.
- Expands documented context capacity from Llama 3.3's 128,000 tokens to one million tokens.
- Introduces sparse mixture-of-experts computation with 17 billion active parameters, rather than Llama 3.3's dense 70-billion-parameter architecture.
Known flaws & limitations
- Can produce inaccurate or fabricated information; important outputs require independent verification.
- Generates text, not native image, audio or video outputs.
- The complete model has approximately 400 billion parameters; 17 billion active parameters do not imply a 17-billion-parameter storage footprint.
- The published benchmark table reports Maverick results, not a separate comparison establishing FP8 parity across serving configurations.
- Deployment is governed by the Llama 4 Community License and Acceptable Use Policy, including restrictions affecting multimodal licensing in the European Union.
Risks, safety & security issues
- Harmful or biased outputs remain possible despite instruction tuning and safety mitigations.
- Adversarial instructions or untrusted retrieved content can challenge application safeguards; tool permissions require independent enforcement.
- Generated text and code can be misused for deception or other prohibited activities.
- Sensitive information in input documents requires deployment-level access controls, retention policies and output handling.
Good fit for
- Image-grounded assistants that return text explanations.
- Long-document analysis and summarization with source checking.
- Multilingual conversational applications in supported languages.
- Self-hosted coding and knowledge assistants with sufficient hardware and human review.
Poor fit for
- High-stakes medical, legal or financial decisions without qualified review.
- Applications requiring native image, audio or video generation.
- Memory-constrained deployments sized for a dense 17-billion-parameter model.
- Autonomous workflows that treat generated instructions as authorization for sensitive actions.
References
Profile last reviewed 10/20/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.