What happened
Meta released Llama 4.5 Scout on July 16 as an open-weight model under Apache 2.0. It targets long-context retrieval and agent planning with a 10-million-token context window at 70B active parameters in a Mixture-of-Experts architecture.
Weights and inference code are on Hugging Face and GitHub, with reference deployments for AWS, GCP and on-prem GPU clusters. Meta claimed Scout matches Gemini 3.5 Pro on several long-document benchmarks while running cheaper on open inference stacks.
Why it matters
Open models with multi-million-token context directly challenge the value of gated frontier APIs for document-heavy workflows. Apache 2.0 licensing removes friction for commercial deployment compared with more restrictive open licenses.
Scout also arrives as enterprises look for ways to cut token spend without giving up long-context capability.
MintedBrain take
Ten-million-token context is only useful if your retrieval and chunking strategy can feed it efficiently. Benchmark on your own document sets and total cost of ownership, including GPU hours, before swapping out a managed API.
Discussion
Sign in to comment. Your account must be at least 1 day old.