Llama 4 Scout and Maverick both activate about 17 billion parameters per token, but that label is not a hardware-sizing shortcut. Scout stores 109B total parameters; Maverick stores 400B. The complete official checkpoint—not the active subset—sets the first memory boundary.
Three numbers that change the hardware plan
- Scout BF16: the official safetensors total 217.284GB. The files exceed 192GB; 240GB leaves only 22.716GB before runtime allocations.
- Maverick FP8: the official safetensors total 416.756GB. Four 96GB accelerators provide only 384GB; five provide 480GB and a 63.244GB file-only margin.
- Maverick BF16: the official safetensors total 803.168GB. Eight 96GB accelerators provide only 768GB; 864GB leaves 60.832GB before runtime allocations.
Those are decimal checkpoint-file totals from Meta’s official repositories. They are not peak runtime memory and do not prove speed. KV cache, image processing, framework workspaces, communication buffers, batch, concurrency and device placement all add requirements.
Why “17B active” is easy to misread
Mixture-of-experts routing means each token uses only part of the network for computation. Standard inference still needs the other experts’ weights available. They may live in accelerator memory, host memory, storage-backed offload or a distributed arrangement, but they do not disappear.
Meta’s statement that Scout fits on one H100 is specifically tied to Int4 quantization. It does not describe the official 217.284GB BF16 checkpoint. Meta’s statement that Maverick can run on one H100 DGX host refers to a multi-accelerator host class, not one 80GB card.
Context is another memory decision
Meta publishes a 10-million-token context ceiling for Scout and 1 million for Maverick. Maximum supported context is not the same as practical context on a minimum-memory deployment. The cache format, actual sequence length, image inputs, batch and concurrency determine the live cost. Hardware planning should begin with the real workload rather than the headline ceiling.
Read the complete capacity guide
The evergreen Llama 4 Scout and Maverick VRAM requirements guide contains the official checkpoint-byte table, file-only boundaries from 192GB through 864GB, context and multimodal caveats, a distributed-deployment checklist and primary sources.
Review multi-GPU placement · Check source-specific measurements · Build a reproducible Llama 4 record · Follow future evidence updates by RSS
Sources
- Meta: The Llama 4 herd
- Meta Llama 4 Maverick official model card
- Official Scout repository API
- Official Maverick FP8 repository API
Published October 25, 2026. File-size arithmetic is not an independent performance benchmark. Vendor deployment statements are labeled as Meta claims.