Hugging Face posted this morning about Tokenmaxxing BERT, a BERT-style tokenizer speed-up that is still sitting in an open pull request on tokenizers v1. AllenAI's October 14 team blog covers HiRO-ACEv2, two Apache checkpoints that take a 100 km climate run down to 3 km storms. FrontiersMind published Lumma-Fev-27B the same day with an official 16.55 GB GGUF. The vLLM Semantic Router's Decision 3.0 card, on the Hub since October 10, entered the default trending list.
On the Hub
Tokenmaxxing BERT
Hugging Face published Tokenmaxxing BERT this morning, a speed-up for BERT-style tokenization that is still sitting in an open pull request on tokenizers v1. The author claims about 50 times the throughput of tokenizers 0.23, figures from the lab rather than an independent bench. The traffic that cares is embeddings and rerankers, not a new chat model, and no new weights landed with the post.1
AllenAI HiRO-ACEv2
AllenAI published a team blog on October 14 for HiRO-ACEv2, a climate emulator rather than a chatbot. Two Apache-2.0 checkpoints take a 100 km global ACE run down to 3 km storms, covering six-hourly precipitation, 10 m winds, and sea-level pressure, and the lab says global 3 km data costs about 1/300th of the X-SHiELD physics model they trained against. Given coarsened X-SHiELD, HiROv2 matches high-resolution statistics to within 3 percent, including events as rare as one in a million, author numbers. The weights landed October 12. The pair only works with the ACE checkpoint tuned to X-SHiELD, so it cannot yet run a warmer future or an ERA5 historical.23
Lumma-Fev-27B
FrontiersMind published Lumma-Fev-27B on October 14 and pushed an official GGUF the same afternoon. It is a full-weight Apache-2.0 fine-tune of Qwen3.8-27B that answers choice and yes-no in one pass, with English plus Hindi, Tamil, and Kannada. The official Q4_K_M file is 16.55 GB. On a held-out set of 4,454 decisions the lab reports 90.7 percent against 83.9 percent for hosted TypeSafe Jev, author-run numbers rather than an independent index.45
vllm-sr/d3
The vLLM Semantic Router team's d3 card climbed onto Hugging Face's default trending list this morning. The 27 billion parameter Apache-2.0 weights have been on the Hub since October 10. Decision 3.0 takes text and images and returns a choice, yes-no, or score, not a paragraph. The author's Jev Decision Index 0.3.2 score is 66.3, first on the official text board.6
