What is Llama (Meta)?
Llama is Meta’s downloadable-weight model family, aimed at developers who want control over inference and adaptation. This entry specifies Llama 4 Scout (17B activated parameters, 16 experts), a text-and-image model with a published 10-million-token context; that headline limit does not mean a small deployment can serve it cheaply.

Key features
- Downloadable model weights
- Mixture-of-experts architecture
- Text and image input; text and code output
- Published 10M-token context
- Twelve supported languages
- Customisation under Meta’s licence
Best for
- Self-hosted internal assistants
- Custom model adaptation
- Multilingual text applications
- Private document analysis
Who it suits
- ML infrastructure teams
- Organisations needing self-managed inference
- Researchers adapting model weights
Who may not need it
- Teams without model-hosting capacity
- Projects requiring an Apache-licensed model
Where it fits in a workflow
Treat Scout as an infrastructure choice, not a ready-made chat subscription. Check licence eligibility and serving-memory requirements before deciding how much context to expose to users.
Specifications
| Model family | Llama |
|---|---|
| Version these details describe | Llama 4 Scout (17Bx16E) |
| Developer | Meta |
| Modalities | text, image |
| Context window | 10,000,000 tokens |
| Access | Downloadable weights, Self-hosted inference |
| Licence | Llama 4 Community License |
Specifications apply to the version named above only and were checked on Oct 26, 2026. Official model documentation.
Sources
Strengths
- Deployment need not depend on a hosted model API
- Weights can be adapted under the licence
- Scout is smaller in total parameters than Maverick
Watch-outs
- 109B total parameters still require substantial hosting resources
- Custom licence is not a permissive open-source licence
- EU-domiciled individuals and companies face multimodal licence restrictions
- Very long context requires infrastructure planning
Alternatives
Mistral models
Open-weight Mistral models for multimodal applications and controlled deployment.
GPT (OpenAI)
OpenAI models for text, image analysis and tool-driven applications.
DeepSeek models
DeepSeek models with long-context reasoning and compatible developer APIs.
Qwen
Alibaba’s Qwen family, with text, image and video options for developers.
Related coverage
Cloud vs. On-Premise AI: A Practical Decision Guide
Compare deployment options using workload, data, staffing, and cost—not assumptions about which environment is always safer.
GuideHow to Choose the Right AI Deployment Strategy
Compare managed AI services, self-hosted models, local systems, and hybrid setups by control, cost, reliability, and operational responsibility.
GuideLLM and Multimodal Model Implementation Checklist
Use this launch checklist to move an AI prototype into a monitored service with defined tests, data controls, and fallback behavior.
GuideIntroduction to Large Language and Multimodal Models
Learn how language and multimodal models process information, what tokens and context mean, and why plausible answers still need checking.
GuideHow to Choose Between LLM and Multimodal Models
Choose a model by testing the real task, input requirements, response time, and full operating cost—not a leaderboard alone.
GuideWhat Is Open Source Software and How Does It Work?
Learn what open source permits, how community development works, and why free downloads still involve real operating costs.
Listings are editorial and independent. Details that change often are marked “to be verified” until an editor checks them against the official source.