Hugging Face's hf CLI now shares one cached copy of a file across repos, and its own benchmark says one in three CLI users is an AI agent
AspireAI community · Hugging Face Blog — huggingface
The huggingface_hub and hf CLI releases cover shared cache, Jobs, Storage Buckets, Sandboxes and download improvements. The cache now stores each Xet file once, so repos sharing a file such as the 9.5 GB text encoder shipped in FLUX.1-dev, FLUX.1-schnell and FLUX.1-Krea-dev get a link instead of a second download, and hf cache prune removes shared files no repo uses. Jobs gained retries with hf jobs run --attempts, reruns with the same secrets and hardware, schedules, port exposure and webhooks that can watch Storage Buckets and pass secrets, plus a UV script config carried in the script header. Downloading from a Storage Bucket is much faster: HfFileSystem now downloads Xet files straight to disk, and the authors report a 2.15 GB parquet file going from about two minutes to about seven seconds on HF Jobs, and the same file through the datasets library from 130 seconds to about 17 seconds with datasets 5.1.0. Sandboxes, still experimental, run code in isolated machines on Jobs, and the import is 6 times faster at 0.39 seconds against 2.36. The most useful number for an agent builder is the skill benchmark: on a given day one in three users of the hf CLI is an AI agent, and more than half of the users uploading to the Hub with the CLI do it through an agent. In their test with DeepSeek-V4.1-Flash and Kimi-K3 on 28 Hub tasks, both models solved 99 to 100 percent of tasks either way, but with the CLI skill they needed about a third fewer turns.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.