Advertised context window in tokens; 17B activated parameters across 16 experts.
Llama 4
Decoded.
An executive-ready, source-linked analysis of Meta's multimodal mixture-of-experts models: Scout and Maverick. Architecture, capability, measured results, deployment considerations and responsible adoption.
Two models. Different deployment priorities.
Experts in the MoE architecture, with 17B activated parameters and 1M-token context.
Multilingual text and image input; multilingual text and code output, not native image output.
Scout vs. Maverick
| Attribute | Llama 4 Scout | Llama 4 Maverick |
|---|---|---|
| Architecture | Autoregressive, native multimodal MoE | Autoregressive, native multimodal MoE |
| Activated parameters | 17B | 17B |
| Total parameters | 109B | 400B |
| Experts | 16 | 128 |
| Advertised context | 10 million tokens | 1 million tokens |
| Pretraining tokens | ~40 trillion | ~22 trillion |
| Knowledge cutoff | August 2024 | August 2024 |
| Modalities | Text + images in; text + code out | Text + images in; text + code out |
| Weights / deployment | BF16; Meta describes single-H100 fit with on-the-fly INT4 quantization | BF16 and FP8; Meta describes FP8 fit on one H100 DGX host |
Source: Meta model information and quantization sections. Context-window maximum does not imply consistent accuracy throughout the full window or affordable production inference.
Native multimodality
Early fusion combines visual and language processing. Intended tasks include visual question answering, image recognition, chart and document understanding, and image reasoning.
Language coverage
Meta explicitly lists 12 supported languages: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. Turkish is not in the explicitly supported list; validate separately before Turkish-language production use.
Instruction-tuned benchmark explorer
Reported Meta evaluations on BF16 models. Higher scores are better within each individual test; scores across different tests are not directly comparable.
MMLU Pro — reasoning and knowledge
Reported scores; 0-shot, metric: macro average accuracy.
Deployment decision framework
01. Use-case fit
Define whether long-document analysis, visual reasoning, coding, multilingual support or interactive assistant response is the main requirement.
02. Infrastructure fit
Benchmark the actual checkpoint, quantization, memory footprint, throughput and cost per successful task on the intended hardware.
03. Risk controls
Test hallucinations, prompt injection, sensitive data leakage and misuse. Meta describes Llama Guard, Prompt Guard and Code Shield as complementary system protections.
Potential ImageFirm pilots
- Multimodal document and presentation analysis
- Long-context internal research assistant with retrieval and citations
- Visual extraction from charts and marketing materials
- Code assistance with human review and automated testing
Illustrative pilot opportunities, not claims of demonstrated model performance.
Governance gates
- Review Llama 4 Community License and Acceptable Use Policy
- Check data protection, retention and vendor architecture
- Validate language and image-count limitations
- Require scenario-based evaluation before launch
Meta reports image-understanding testing up to five input images; additional-image use needs developer validation.
Training footprint & limitations
Combined H100-80GB GPU hours reported for pretraining.
Estimated location-based CO₂e for training across both models.
Reported market-based CO₂e under Meta's renewable-energy accounting; this does not mean physical emissions were zero.
Meta notes that model outputs can be inaccurate or objectionable and that no evaluation can cover every deployment scenario. The model card describes static models trained on offline datasets; neither model provides current knowledge without an external data system.
What the numbers do—and do not—say
Vendor-reported ≠ independently reproduced
The benchmark explorer reproduces instruction-tuned scores published in Meta’s Llama 4 model card. No ImageFirm evaluation, third-party replication or current frontier-model comparison is claimed.
Source: Meta model card, “Benchmarks”, instruction-tuned models.
Context length ≠ assured recall
The published 10M and 1M token limits describe supported context capacities. Retrieval accuracy, throughput, total cost, latency and multilingual performance require task-specific testing at the intended context lengths.
ImageFirm implementation guidance; not a measured result.
Language and licensing boundaries
Meta lists 12 officially supported languages. Other languages may require additional training, testing and risk mitigation. Commercial adoption must be assessed against the Llama 4 Community License and Acceptable Use Policy.
Recommended pilot scorecard
Track answer correctness, evidence traceability, prompt-injection resistance, p50/p95 latency, tokens per request, cost per successful task, and human escalation rate. Define acceptance thresholds before evaluating models.
ImageFirm proposed evaluation design.
Primary references & further reading
Official source links are provided for verification. This page is an independent ImageFirm synthesis, not an official Meta publication or endorsement.
Facts and reported benchmark values are sourced from [1]. Commentary, selection guidance and pilot suggestions are ImageFirm analysis. Specifications and external pages may change; verify before procurement or deployment.