Llama 4 Scout 17B-16E Instruct

Checkpoint identity

Official identifier

meta-llama/Llama-4-Scout-17B-16E-InstructMeta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Architecture

Autoregressive mixture-of-experts model with early-fusion native image/text inputMeta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Total parameters

109 billionPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Meta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Active parameters

17 billion per tokenActive parameters are a compute-related count, not the model's full stored weight count.Meta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Modalities

text-input, image-input, text-outputMeta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Context

Not verified for this recordNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Meta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Release date

2025-04-05Meta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Download access

Gated: publisher terms / access acceptanceMeta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

License

Llama 4 Community LicenseA custom commercial license governs the checkpoint, with attribution, redistribution, derivative naming and acceptable-use obligations. A release-date monthly-active-user threshold greater than 700 million requires a separate Meta grant. Weight download requires accepting the publisher's access conditions; listing the model does not accept those terms on Aguasto's behalf.Meta Llama4Scout17B16EInstruct official model card (external source) · Llama4Community License (external source)

Theoretical weight-memory footprint

Using the publisher's approximately 109 billion total parameters at 16 bits per parameter gives 218 GB (203.0 GiB) of nominal weights.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

The model card's Transformers example loads BF16 and discusses quantization as a separate deployment choice. Its conditional hardware example does not establish a full-context or production-fit result. Aguasto has not downloaded or tested gated weights.

Serving considerations

• The publisher distinguishes 109B total from 17B activated parameters. Expert routing reduces activated work per token; it does not make the full checkpoint a 17B weight-storage problem.
• The model card advertises 10M context, but the exact numeric checkpoint position limit was not readable without gated access. The numeric context field stays unknown rather than turning the rounded advertisement into an exact config value.
• Native early-fusion text/image inputs require the correct processor and chat template. The publisher describes testing image understanding up to five images; additional-image behavior needs a separate evaluation.
• The card provides a Transformers implementation path. Model-specific vLLM/SGLang compatibility is not asserted by this record; supported model architecture, version, attention and parallelism must be checked independently.
• No Aguasto inference offering, deployment, benchmark, GPU fit or throughput is claimed. The record is a technical research reference for this exact checkpoint.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.