The Hub is quieter than usual. Four updates landed overnight. They cover OCR, a zero token judge, a DNA annotator, and a fine tune library change.
On the Hub
LightOnOCR-3
LightOn published a Hugging Face blog for LightOnOCR-3. It is one model for transcription, layout grounding, image descriptions and chart extraction. Three sizes are on the Hub under Apache-2.0. The 4B card is from October 6. The 1B and 0.8B cards are from October 7.1
Darwin-27B-ZTC
FINAL-Bench added Darwin-27B-ZTC to the Hub on October 8 and published a blog later that morning. It is a 27B zero token classifier that returns calibrated probabilities in one forward pass with no generated tokens. The card is Apache-2.0. The author claims Typed Decisions accuracy of 0.743. This is not the September Darwin-180B-RSI card.2
Carbon-A
Hugging Face Biology Research published the Carbon-A blog on October 8. The Hub card HuggingFaceBio/Carbon-A-1.2B has been up since October 2 under MIT. It is a 1.2B DNA annotator with 6-mer tokens and a 98,304 bp window. It predicts protein coding sequence at single base resolution on both strands.3
TRL v1.15.0
Hugging Face tagged TRL v1.15.0 on October 8. SFT, DPO, KTO, GRPO, RLOO and Distillation now score with a fused LM head by default. The published gemma-3-1b table shows KTO max sequence up 6.89x. At 8,192 tokens peak memory is 52 to 82 percent lower. Triton is required for the fused path.4
