Llama
The open-weight family powering thousands of independent projects.
Olya's profile
- What sets it apart
- Llama isn't a service but a family of models with downloadable weights: you run them yourself, on your hardware or through your provider, without going through a Meta API. Mind the word "open": it's open-weight, not open source in the strict sense, and the license requires a separate agreement for anyone above 700 million monthly users. It's that availability, not any single benchmark, that made it the de facto standard for fine-tuning and independent projects.
- Strengths
- The latest generation, Llama 4 Scout and Maverick (April 2025), uses a mixture-of-experts architecture: 17 billion active parameters per token out of 109 and 400 billion total, so strong output at modest compute cost, with Scout fitting on a single H100 GPU when quantized. They're natively multimodal (text and images) and Scout claims a context window of up to 10 million tokens. Around it sits the widest ecosystem out there: local tools, tuned variants, and communities covering every use case.
- When to use it
- It pays off when you want control: data that stays in-house, on-premise or self-hosted execution, no lock-in to a single vendor. It's the natural pick for fine-tuning on a specific domain or building a product without paying per call. It's also great just for learning, since you can inspect and adapt the model.
- When to avoid it
- If you'd rather not manage infrastructure, a closed, ready-made API gets you to a result faster. The giant context windows are a headline figure: real performance on long documents has to be tested on your own case, not assumed. And keep an eye on Meta's commitment to open: no new Llama arrived in 2026 and the company pivoted to a closed-weight model, a sign the road ahead is less settled than it looks.