HuggingFace Gets Service Accounts for CI/CD, a Hardware-Fit Model Filter, and Transformers v5.13.0

TL;DR: HuggingFace ships service accounts for org CI/CD (no more person-tied tokens), a hardware-fit model filter, and Transformers v5.13.0 with Kimi 2.5, Nemotron 3.5 ASR, and unified HfExporter.

Service accounts for enterprise orgs (July 3): Enterprise Hub organizations can now create service accounts — organization-owned identities for automation, CI/CD, and backend integrations that aren't tied to any individual. Automated pipelines and token rotation no longer break when team members leave or change roles; service accounts show up in audit logs under the org rather than a person's account. This closes the most common ML-team CI/CD anti-pattern.

Hardware filter on Models page: A new Hardware filter on the HuggingFace Models page lets you filter to models that fit your specific GPU, CPU, or Apple Silicon chip. If you're picking a local model for Ollama or vLLM, you now only see what will actually run on your machine before downloading.

Transformers v5.13.0: Adds Kimi 2.5 (Moonshot AI), MiMo-V2-Flash, and Nemotron 3.5 ASR; unified HfExporter for PyTorch, ONNX, and ExecuTorch in one call; and a batch of generation, quantization, cache, and attention fixes.

Why it matters: Service accounts finally close the "CI token tied to a person" anti-pattern that causes silent breakage in every ML team's automation stack.

Sources: HuggingFace blog · HuggingFace Releasebot — July 2026