LLaMA 4 Maverick is a cutting-edge multimodal model featuring 17 billion active parameters within a Mixture-of-Experts architecture of 128 experts, totaling 400 billion parameters. It leads its class by outperforming models like GPT-4o and Gemini 2.0 Flash across a wide range of benchmarks, and it matches DeepSeek V3 in reasoning and coding tasks—using less than half the active parameters. Designed for efficiency and scalability, Maverick delivers a best-in-class performance-to-cost ratio, with an experimental chat variant achieving an ELO score of 1417 on LMArena. Despite its scale, it runs on a single NVIDIA H100 host, ensuring simple and practical deployment.
Llama 3.3 70B Instruct, created by Meta, is a multilingual large language model specifically fine-tuned for instruction-based tasks and optimized for conversational applications. It is capable of processing and generating text in multiple languages, with a context window supporting up to 128,000 tokens. Launched on December 6, 2024, the model surpasses numerous open-source and proprietary chat models in various industry benchmarks. It utilizes Grouped-Query Attention (GQA) to improve scalability and has been trained on a diverse dataset comprising over 15 trillion tokens from publicly available sources. The model's knowledge is current up to December 2023.
| Llama 4 Maverick | Llama 3.3 70B Instruct | |
|---|---|---|
Web Site
| ||
Provider
| ||
Chat
| ||
Release Date
| ||
Modalities
| text images video | text |
API Providers
| Meta AI, Hugging Face, Fireworks, Together, DeepInfra | Fireworks, Together, DeepInfra, Hyperbolic |
Knowledge Cut-off Date
| 2024-08 | 12.2024 |
Open Source
| Yes (Source) | Yes |
Pricing Input
| Not available | $0.23 per million tokens |
Pricing Output
| Not available | $0.40 per million tokens |
MMLU
| Not available | 86% 0-shot, CoT Source |
MMLU-Pro
| 80.5% Source | 68.9% 5-shot, CoT Source |
MMMU
| 73.4% Source | Not available |
HellaSwag
| Not available | Not available |
HumanEval
| Not available | 88.4% pass@1 Source |
MATH
| Not available | 77% 0-shot, CoT Source |
GPQA
| 69.8% Diamond Source | 50.5% 0-shot, CoT Source |
IFEval
| Not available | 92.1% Source |
SimpleQA
| - | - |
AIME 2024 | - | - |
AIME 2025 | - | - |
Aider Polyglot
| - | - |
LiveCodeBench v5
| - | - |
Global MMLU (Lite)
| - | - |
MathVista
| - | - |
Mobile Application | - | - |
VideoGameBench | ||
Total score | 0% | - |
Doom II | 0% | - |
Dream DX | 0% | - |
Awakening DX | 0% | - |
Civilization I | 0% | - |
Pokemon Crystal | 0% | - |
The Need for Speed | 0% | - |
The Incredible Machine | 0% | - |
Secret Game 1 | %0 | - |
Secret Game 2 | 0% | - |
Secret Game 3 | 0% | - |
Compare AI. Test. Benchmarks. Mobile Apps Chatbots, Sketch
Copyright © 2025 All Right Reserved.