Llama 4
Meta's natively multimodal open MoE herd with industry-leading context length.
Context window
Up to 10M tokens (Scout); ~1M tokens (Maverick)
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Meta
Open-weight models under the Llama 4 Community License (commercial use permitted for organizations under 700M monthly active users; larger orgs must request a license). Free to download; running costs depend on your hardware or inference provider pricing (e.g., Together AI, Fireworks, Hugging Face). · Data verified 2026-07-02
Llama 4 is Meta's first herd of natively multimodal, open-weight models, released April 5, 2025. It uses a mixture-of-experts (MoE) architecture with early-fusion multimodality. Llama 4 Scout has 17B active parameters (109B total) across 16 experts and an industry-leading 10M-token context window, fitting on a single H100 with int4 quantization. Llama 4 Maverick has 17B active parameters (400B total) across 128 experts and supports up to ~1M tokens, positioned as a top multimodal model in its class. A larger Behemoth teacher model (~288B active / ~2T total) was previewed but not released. Weights ship under the Llama 4 Community License.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Download open weights from Hugging Face under the meta-llama organization (Scout and Maverick, base and instruct) or from llama.com. Run via Transformers (v4.51.0+), vLLM, or hosted providers. Scout fits a single H100 with int4 quantization.
Strengths
- ✓Open weights with native multimodality (early fusion)
- ✓Industry-leading context length (up to 10M tokens on Scout)
- ✓Efficient MoE design - only 17B active parameters
- ✓Scout fits a single H100 GPU with int4 quantization
- ✓Full Hugging Face / Transformers integration for fine-tuning
Best for developers who...
When to choose it (and when not to)
Reach for Llama 4 when...
- →You need extremely long context in an open model (Scout's 10M window)
- →Self-hosted or on-prem multimodal deployment
- →You want an efficient MoE that activates few parameters per token
- →Fine-tuning or full control over the model
Look elsewhere if...
- ✕You are a very large org (700M+ MAU) unwilling to request a Meta license
- ✕You want a fully managed API with no infrastructure
- ✕You need the newest frontier reasoning quality (consider frontier hosted models)
- ✕Limited GPU memory for the larger Maverick variant
How to use it
- ›Use the instruction-tuned checkpoints for chat and assistant tasks
- ›Apply Llama 4's chat template and system prompt conventions
- ›For very long context, feed full documents/repos rather than chunking
- ›Quantize (int4/FP8) to fit available GPUs - Scout int4 fits one H100
Quickstart
Pythonfrom transformers import pipeline
pipe = pipeline("text-generation", model="meta-llama/Llama-4-Scout-17B-16E-Instruct", device_map="auto")
messages = [{"role": "user", "content": "Summarize the plot of Hamlet."}]
print(pipe(messages, max_new_tokens=256)[0]["generated_text"])Requires transformers >= 4.51.0 and accelerate. Accept the Llama 4 license on Hugging Face and run `huggingface-cli login`. Also available on Together AI and Fireworks.
API model id: meta-llama/Llama-4-Scout-17B-16E-Instruct
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Scout context window | 10M tokens | Longest context of any open or closed model at release |
| Scout size | 17B active / 109B total (16 experts) | MoE, natively multimodal |
| Maverick size | 17B active / 400B total (128 experts) | Beats GPT-4o and Gemini 2.0 Flash per Meta |
Source: Meta AI - The Llama 4 herd
Compare Llama 4 with
Llama 4 vs DeepSeek V3
DeepSeek - 128K tokens ctx
Llama 4 vs Qwen 3
Alibaba (Qwen Team) - 128K tokens (32K for 0.6B/1.7B/4B dense variants) ctx
Llama 4 vs Gemma 3
Google DeepMind - 128K tokens (32K for the 1B variant) ctx
Llama 4 vs Mistral Large
Mistral AI - 128000 ctx