Hugging Face is putting numbers on text to speech. The audio team published the Open TTS Leaderboard blog on Sep 30 and the Space has been live since Sep 7. A CI agent that ships Transformers fixes and a materials screening contest are also moving on the Hub.
On the Hub
Open TTS Leaderboard
Hugging Face published the Open TTS Leaderboard blog on Sep 30 1. The board adds objective WER, CER, RTFx, TTFA and WavLM speaker similarity for 8,000 plus models on the Hub. The Space hf-audio/open_tts_leaderboard was created Sep 7 and last modified Sep 30 2. Intelligibility checks use Qwen/Qwen3-ASR-1.7B-hf. The board is meant to complement arena voting with reproducible metrics.
Serge
Hugging Face published Anatomy of a bug-fixing agent on Sep 29 3. It describes Serge, a CI agent that landed 29 maintainer merged PRs in 80 days. The loop picks persistent nightly failures, reproduces on GPU, generates a patch with ruff, ty and transformers-mlinter, verifies five runs, then opens a PR. Serge runs in isolated Kubernetes pods with no write access to main. The post frames it as production agentic work on the Transformers repo.
Open Superconductor Challenge
FINAL-Bench published the Open Superconductor Challenge blog on Sep 29 4. The dataset FINAL-Bench/OSC-Superconductor was created Sep 28 with cc-by-nc-4.0 license 5. The leaderboard Space FINAL-Bench/OSC-Leaderboard was created Sep 29 with apache-2.0 license 6. Season 1 closes Dec 31 2026 and offers a $3,000 prize pool for screening 2D materials for d-wave pairing on a laptop CPU. The blog invites agents to participate and notes the index is a tendency, not a measured Tc.
