Keep the format dumb: ten years of Hugging Face, measured from its own releases

Reads a decade of one company's decisions as a single argument about where a capability belongs, in the file format every reader must parse or in the loader only you control. A reader leaves able to defend that boundary in a design review with dated evidence on both sides, including the counter-experiment where a rival format absorbed quantization and then had to remove two of its types, and able to measure a dependency's architectural history from its own wheels and registry metadata in about twenty minutes.

What you get out of it

• The durable asset across ten years is not a framework or a service but an interchange format that refuses capability: no compression, no code, no holes in the buffer, a 100 MB header cap, and roughly 400 lines of implementation against the 210,000 of the alternative its own README rejects.
• Two closed pull requests state the rule more clearly than any documentation: quantization was refused because no encoding had converged, and a GPUDirect Storage path was refused because 'GDS should be an implementation detail, not something users should care about'.
• GGUF ran the counter-experiment and paid for it. It put quantization types in the specification, and its type enum now carries the receipt, 'GGML_TYPE_Q4_2 = 4, support has been removed'.
• The decision to drop TensorFlow and Flax was legible in the artefacts fifteen months before it was taken: those file counts sat at exactly 73 and 40 in every release measured between October 2024 and October 2025, then went to zero in one step.
• Removing pickle from the weights file relocated code execution rather than ending it: a 2026 advisory shows a model's own config.json supplying the trust_remote_code value that overrode an explicit refusal, and a second shows a serving layer hardcoding the flag and deleting the user's ability to say no.

Scope

Why this, now. Transformers v5 shipped on 26 January 2026 and deleted two entire framework backends while leaving the remote-code switch untouched, and every transformers advisory published since May 2026 is an attack on that switch or on something downstream that hardcoded it.

What it does not cover. Training, the Hub's server side, GPU and inference economics, and every figure that exists only on huggingface.co, including achieved deduplication ratios, bytes stored and cost, because this session's network policy blocked every engineering-blog, video, paper and vendor host, so the guide contains no blog, talk or paper sources at all.