Download the right model files
Distinguish full models, shards, adapters, projectors, and calibration artifacts.
Reviewed 2026-09-08 · Beginner
A repository is a package, not one weight file
A repository can contain several quantizations, tokenizers, configuration files, documentation, and auxiliary components. Its total storage is not necessarily the memory needed for one selected model.
A safetensors checkpoint often spans several shards referenced by an index. A sharded GGUF variant also needs its complete matching shard set. One shard is not the full model, even if its name looks plausible.
Know what is not a standalone model
A LoRA adapter requires a compatible base. A vision projector is a support component. An importance matrix or calibration file helps create a quantization; it is not a complete quantized model.
Draft models used for speculative decoding may need a separate target model. Budget the pair according to the runtime rather than treating the draft’s small size as the whole service.
Keep a reproducible manifest
Record the owner/repository, revision, exact artifact name, size, and runtime version. Check the publisher’s declared base relationship instead of guessing from similar names.
Avoid mixing shards from different revisions. If a file list is incomplete or a base relationship cannot be verified, treat the resulting memory figure as unavailable until the missing evidence is resolved.