Core capabilities
- Native text-image multimodality with early fusion
- 10M maximum context for large corpora and codebases
- 109B total / 17B active MoE architecture
- Open weights available through Meta and partners
← Model indexMODEL PROFILE / llama-4-scout
Meta / open
Llama 4 Scout
A natively multimodal open-weight MoE model designed for single-H100 efficiency and exceptionally long context.
Context10M
ModalityText · Image
AccessWeights · Partners
ReasoningGeneral
Where it fits.
Distinctive for controlled deployments that combine very long inputs with image-text understanding. Scout is a natively multimodal MoE with 109B total, 17B active parameters and 16 experts. Meta highlights a 10M context and single-H100 deployment with Int4 quantization.
Best-fit workflows
- Ultra-long-context analysis
- Private multimodal deployment
- Customization and distillation
04 / OPERATIONSFrom evaluation
From evaluation
to production.
Deployment notes
Use supported partner stacks or a tested self-host runtime. Int4 can fit on one H100 according to Meta, but production memory also includes KV cache and concurrency; benchmark target sequence lengths, not only empty-model fit.
Watch before choosing
Practical throughput, memory use and retrieval quality at 10M context depend heavily on serving and quantization. The published 10M capacity is not a guarantee of uniform recall at every position. Long-context systems still need retrieval tests, positional probes, memory planning and careful image-count validation.
05
Family & versions
- Llama 4 ScoutLong-context efficient
- Llama 4 MaverickHigher capability
Specifications and positioning checked against provider documentation.
Open official source↗