DEV Community

Lingdas1
Lingdas1

Posted on • Originally published at github.com

Llama 4: Meta's Latest — Scout, Maverick, and the MoE Revolution

#ai #opensource #llm #tutorial

Llama 4: Meta's Latest — Scout, Maverick, and the MoE Revolution

The open-source default just got a massive upgrade. Here's what's new and which variant you should actually use.

Llama 4 at a Glance

Meta released Llama 4 in April 2025 with a fundamental architecture change: Mixture of Experts (MoE). Two variants were launched simultaneously:

Variant Architecture Total Params Active per Token Min VRAM (Q4)
Llama 4 Scout 17B × 16 experts 109B ~17B 10 GB
Llama 4 Maverick 17B × 128 experts 2T ~17B 10 GB

Both are available on Ollama as llama4:latest (points to Scout) and llama4:maverick.

💡 The story that sells itself: Meta spent millions training a 2-trillion-parameter model and you can run it on a used gaming GPU. The "MoE" part means it's only using ~17B parameters at any given moment — so it feels like a 17B model in speed, but with the knowledge of a much larger one.

Quick Start 

# Scout (balanced — good default)
ollama pull llama4:latest

# Maverick (bigger knowledge, same speed)
ollama pull llama4:maverick
Enter fullscreen mode Exit fullscreen mode

⚠️ Verify before pulling: Model names on Ollama change. Check https://ollama.com/library/llama4 for current tags.

Scout vs Maverick: Which One? 

Your use case?
├── General chat, writing, everyday coding → Scout (llama4:latest)
├── Deep knowledge, fact-heavy tasks, research → Maverick (llama4:maverick)
├── Speed-critical, low VRAM → Scout
└── Both run at the same speed per token — the difference is knowledge breadth
Enter fullscreen mode Exit fullscreen mode

The practical difference: Maverick has 128 experts vs Scout's 16. This means Maverick's "collective knowledge" is much broader — it's seen more patterns, more facts, more edge cases. But per-token speed is nearly identical because both only activate ~17B parameters at a time.

For most people: start with Scout, upgrade to Maverick if you need more depth.

What Llama 4 Excels At

Task Rating Notes
General conversation ⭐⭐⭐⭐⭐ Natural, helpful, rarely hallucinates
Creative writing ⭐⭐⭐⭐ Good, but Claude-level models still edge it out
Coding ⭐⭐⭐⭐ Strong general coding, weaker at math-heavy tasks
Multilingual ⭐⭐⭐⭐ Supports 8 languages natively
Long context ⭐⭐⭐ 128K context works but quality degrades past 64K

The "But Meta Says I Can't Use It Commercially" Issue

This comes up constantly. Here's the actual situation as of May 2026:

  • Llama 4 is NOT the old "Llama 2 Community License" — it's under the Llama 4 Community License, which is significantly more permissive
  • Commercial use is allowed for companies under 700 million monthly active users
  • You can fine-tune and distribute your fine-tuned versions
  • The license restricts using Llama outputs to train competing models

For indie developers, startups, and small businesses: you're free to use it commercially. For FAANG-sized companies: you need a separate agreement with Meta.

If you want truly unrestricted open-source, use DeepSeek-R1 (MIT) or Qwen (Apache 2.0).

Real-World Benchmarks (Community-Tested)

On an RTX 4090 (24GB):

Model (Q4_K_M) tok/s MMLU-Pro HumanEval
Llama 4 Scout ~45 68.2 76.8
Llama 4 Maverick ~42 72.1 79.3
DeepSeek-R1 32B ~22 74.5 84.1
Qwen 3.6 32B ~25 73.0 81.4

Takeaway: Llama 4 Scout/Maverick are the fastest high-quality models you can run locally. If speed matters more than raw benchmark scores, they're the pragmatic choice.

Pro Tips

  1. Use llama4:maverick with a 32K context limit — the full 128K eats VRAM and degrades attention quality
  2. Don't use Q2/Q3 quants — MoE models lose coherence more sharply at extreme quantization than dense models
  3. Scout is the sweet spot for most setups — unless you're doing research or fact-heavy work

Related guides: Gemma 4 | Qwen | MoE Models

Top comments (0)