Hugging Face Daily Digest Sep 12, 2026

The Hub is quieter than yesterday. Two new GGUF packs appeared overnight. Hugging Face shipped two library tags on Sep 10.

On the Hub

bartowski Swift-Qwen3.8-27b GGUF

bartowski published the first imatrix GGUF pack of UkisAI Swift-Qwen3.8-27b this morning. The pack uses llama.cpp b10896 with 42 layout files. The card recommends Q4_K_M at 17.44 GB. The license is Swift Open License 1.0 with a US$1,000,000 annual recurring revenue cap. ^{1 2}

bartowski Scarlet-Shadow-31B GGUF

bartowski published the first imatrix pack of Vortex5 Scarlet-Shadow-31B this morning. The pack is Apache-2.0 and built on llama.cpp b10896 with 42 layout files. The recommended Q4_K_M is 19.52 GB. The base is Vortex5/Scarlet-Shadow-31B, a Gemma-4 31B merge. ^{3 4}

TRL v1.13.0

Hugging Face tagged TRL v1.13.0 on 2026-09-10. The release removes PPOTrainer, PPOConfig, and the value head code. Chunked cross entropy for lm_head stays in bf16 on tensor cores. The notes report a 6.0x speed up for one chunk on an H100. ⁵

huggingface_hub v1.31.0

huggingface_hub v1.31.0 was tagged on 2026-09-10. The release adds custom labels for dedicated sandboxes. snapshot_download now writes refs atomically to avoid races. HfFileSystem.get validates remote filenames before local write to block path traversal on Windows. ⁶