Llama 4
Llama 4
Llama 4 is Meta's fourth-generation open-source LLM family, released in 2025-2026. It includes multiple scales and multimodal capabilities.
Model Variants
| Variant | Parameters | Description |
|---|---|---|
| Llama 4 Scout | 17B | Lightweight, fastest |
| Llama 4 Maverick | 17B MoE | Balanced quality/speed |
| Llama 4 Behemoth | 288B MoE | Largest, best quality (research only) |
Key Specifications
- Context window: 128K tokens (Scout, Maverick), 256K (Behemoth)
- Max output tokens: 8,192
- License: Custom (open-source for most use cases)
- Architecture: Transformer with GQA
Architecture
Llama 4 uses a dense transformer for Scout, MoE for Maverick and Behemoth. All use RoPE positional encoding, GQA, and pretraining on large-scale web data.
Capabilities
- Open weights — Available for download and self-hosting
- Fine-tuning — Supports LoRA and full fine-tuning
- Ecosystem — Broad tooling support (llama.cpp, vLLM, Ollama, etc.)
- Multimodal — Vision capabilities via integration
Deployment
Llama models can run locally via local deployment with llama.cpp, Ollama, or vLLM. The 17B Scout variant runs on consumer GPUs with quantization.
Comparison
- DeepSeek V3 — Open-source competitor
- Mistral Large — Open-source competitor
- qwen-2.5 — Open-source competitor
- See Open-Source vs Closed-Source LLMs for the open vs closed model landscape