Model Weights

Open weights differ from fully open source: the former makes model parameters public but does not necessarily disclose training data and complete training code.

Open-source models generally refer to AI models whose model weights can be publicly downloaded and which allow free use and modification, as opposed to closed-source models that are only accessible via API.

Model weights are the parameter values stored after a neural network has been trained; releasing weights publicly means anyone can run, fine-tune, and deeply audit the model locally.

The weight file of a 7-billion-parameter language model occupies approximately 14 GB of storage, while training data often reaches several TB or even tens of TB.

If hardware is physically obtained, attackers may bypass encryption through side-channel methods such as differential power analysis and electromagnetic emanation analysis.

When deploying large models locally, VRAM usage consists of two parts: model weights and runtime memory (primarily the KV cache)

Model weights refer to the set of parameters in a neural network obtained through training; a weight refresh means replacing the original parameters with newly trained or fine-tuned parameters without changing the model's external name or calling interface.

Security teams typically audit an agent's code, dependencies and API permissions, but rarely audit the weight provenance of the underlying model; fine-tuned models from unofficial channels or 'unrestricted' variants shared by the community may carry abliteration processing, which is especially hard to detect in model registries and third-party model marketplaces.

Model weights should be brought into the scope of software bill of materials (SBOM) management, clarifying which model versions have passed security assessment and which are prohibited from use in production agents.

Even with completely identical model weights, the actual serving cost and response speed can differ severalfold without engineering optimizations such as inference optimization and concurrent scheduling.

Model weights are merely parameter files saved after a neural network has converged in training and do not themselves contain runtime logic.

Model weights are only part of VRAM usage; during inference, space must also be reserved for activations, KV cache, context window, and so on.

The draft's scope covers model weights, source code, and training methodology of systems meeting the above threshold.

Listing model weights themselves as export-control items is uncommon in export-control practice.

Ways of transferring weights include directly downloading files, distributing via container images, and embedding them in applications for distribution; the draft's inclusion of weights under control means that these actions may trigger licensing requirements under specific circumstances.

The draft for comments stipulates that restricted underlying large-model technologies include model weights, source code, and architecture.

Model weights are parameter files fixed after training (such as .gguf and .safetensors) that only determine how the model answers and do not store user conversations.

Manually deleting model weights or vector indexes will cause the tool to re-download or rebuild them on next startup, which is time-consuming and may incur traffic costs.

The company stated that there is currently no indication that the employee improperly accessed or stole model weights, source code, or customer data.

Model weights are the large set of parameters obtained after a neural network is trained; they determine the model's behavior and capabilities and are an AI company's most core intellectual property.