The Hub is quiet on big lab drops. Hugging Face shipped a transformers update for Apple Silicon. Base Compute opened a public local inference leaderboard. Hugging Face also published a repository memory tool for coding agents.
On the Hub
Transformers now loads GGUF on Apple Silicon
Hugging Face posted an update on September 22.1 Transformers can now load Hub GGUF files with from_pretrained. The work uses ggml Metal kernels via the kernels library. Early focus is Apple Silicon and Qwen3.5 models. The blog shows unsloth/Qwen3.5-4B-GGUF in Q4_K_M and a serve example with transformers serve.
Compute:Arena opens a public local inference leaderboard
Base Compute published a community blog on September 22.2 It open sourced Compute:Arena, with the site at computearena.ai and the CLI under Apache-2.0. It benchmarks model times, quant levels and chip runtime. Each run signs a local report with hashes and telemetry. The leaderboard covers BaseRT and llama.cpp llama-bench.
relore adds repository memory for coding agents
Hugging Face published a community blog on September 21.3 It introduces relore, a tool that indexes git history for coding agents. It ranks maintainer decisions over contributor guesses. Queries run locally and ingestion contacts GitHub. Code is at github.com/huggingface/relore and the license is Apache-2.0.
