G3
Google DeepMind
Gemma 3 27B IT
Gemma 3 27B IT is Google's 27B multimodal instruction model with 128K context, text and image input, 140+…
Free
L4
Meta Advanced
Llama 4 Scout
Llama 4 Scout is Meta's 17B-active/109B-total multimodal MoE with a 10M-token context window, multilingual support and downloadable weights.
Open-Weight Language ModelTextImage Free
What is this model and why does it matter?
Llama 4 Scout is Meta's 109B-total mixture-of-experts model with 17B active parameters and an unusually large 10M-token context window. It accepts multilingual text and images.
Very long-context researchMultimodal assistantsSelf-hosted AIFine-tuningPrivate deployments
Llama 4 Scout: features, use cases and important details
Llama 4 Scout is Meta’s long-context multimodal model from the Llama 4 family.
Llama 4 Scout verified specifications
Meta’s official model card lists 17B active parameters, 109B total parameters and a 10-million-token context window. Scout accepts multilingual text and images and generates multilingual text and code.
License and deployment
The weights are distributed under the Llama 4 Community License and can be deployed on private infrastructure or through cloud providers. There is no single Meta token price for the downloadable weights.
Best uses
Scout is particularly attractive for massive document collections, multimodal research, private assistants and custom deployments that benefit from very long context.
Limitations
It has an August 2024 knowledge cutoff, substantial hardware requirements and a custom community license rather than a standard open-source license.
How to use this model
- Review and accept the Llama 4 Community License.
- Download Scout from Meta's official distribution channel or a supported cloud provider.
- Provision enough GPU capacity for the MoE model.
- Use text and image inputs with a compatible serving stack.
- Evaluate accuracy, memory and latency before production.
Example prompts
Analyze this very large document collection and identify contradictions.Compare this image with the attached specification.Summarize this multilingual corpus while preserving key evidence.
What it can do
- 17B active / 109B total MoE
- 10M context
- Text and image input
- Multilingual generation
- Code output
- Downloadable weights
Practical use cases
- Massive-context analysis
- Private assistants
- Research
- Multimodal Q&A
- Custom fine-tuning
What does it cost?
Meta distributes Llama 4 Scout as downloadable weights under the Llama 4 Community License. There is no single Meta token price; hosting cost depends on the infrastructure or cloud provider.
Simple summaryThe weights do not have a Meta per-token charge, but a 109B-total MoE still requires substantial GPU memory and serving infrastructure.
What stands out
- 10M context
- Downloadable weights
- Multimodal input
- Broad language support
- Lower active parameters than total size
Things to consider
- Large deployment footprint
- Custom community license
- Static August 2024 knowledge cutoff
Important restrictions and trade-offs
- Self-hosting is hardware intensive
- The license is not a standard OSI open-source license
- Long context does not guarantee perfect recall
Our editorial take
One of the most searched open-weight Llama variants because its 10M-token context makes it unusually suited to massive document and multimodal workloads.