Meta introduced Llama 4 Scout and Maverick in April 2025, not June 2026. Meta describes Scout as a mixture-of-experts model with 17 billion active and 109 billion total parameters, and Maverick as having 17 billion active and about 400 billion total parameters. Those figures and benchmark results come from Meta and its model materials.
What open weights change
Open weights allow an organisation or hosting partner to inspect deployment artefacts and run the model in an environment it controls. That can support data-location and operational-control objectives, but it does not automatically establish compliance, security, or fitness for a regulated use.
What still requires proof
Teams must review the applicable licence, establish hardware and quantisation requirements, test accuracy and safety on their own tasks, restrict data access, and monitor the deployed version. Hardware claims vary with precision, context length, batch size, and serving stack, so this review does not repeat a universal GPU-count claim.
Practical conclusion
Treat Scout and Maverick as candidates for a controlled evaluation, not as automatic substitutes for hosted frontier services. Compare them with the same data, quality gates, latency targets, and operational cost model.
Sources & further reading
Follow the original evidence. Sources may include the organisation making the announcement; claims and independent findings are distinguished in the analysis.
01Llama 4 model family, Meta AIai.meta.com02Llama get-started documentation, Meta AIai.meta.com03Llama 4 Scout model card, Hugging Facehuggingface.coLlama 4 Scout and Maverick: The Real Deployment Trade-off for Private AI
Meta released Scout and Maverick in April 2025. Open weights increase control, while licence, hardware, evaluation, and operational requirements remain decisive.
Last factual review: 24 August 2026