The Hub is quiet this morning, and four items still set the day. LightOn wrote up a 2B local code-search agent with an official 2.7 GB GGUF, and AutoTrust added a 35B-A3B decision model whose NVFP4 pack fits a 32 to 48 GB card. OpenBMB's MiniCPM-V line now has a 35B MoE whose official GGUF was touched this morning. DataScience-UIBK published a 179M retriever that keeps a dense head and a ColBERT head in one MaxSim.
On the Hub
LightOn ColGrep-Agent-2B
LightOn published a Hub post on October 12 for ColGrep-Agent-2B, an Apache-2.0 2B MiniCPM5-2B fine-tune that sits between a coding agent and the repo. It searches with ColGrep, reads files through a read-only terminal, and returns paths plus line ranges. The official GGUF is 2.68 GB at Q8_0, with a 1.56 GB Q4_K_M beside it. LightOn's own table, against Claude Code plus Sonnet 5.5 grep on 52 code-search questions, says 34 percent fewer API tokens and no real accuracy drop on those search questions. The weights have been on the Hub since October 3, so the news is the write-up and the measured cut.12
AutoTrust JEV-35B
AutoTrust put JEV-35B on the Hub on October 11, a new Apache-2.0 size in the JEV decision-model line: a Qwen3.5-35B-A3B MoE with about 3B active, built to pick among up to 256 options in one pass. The useful local fact is the NVFP4 sibling, 25.6 GB to download and 23.3 GiB on the GPU, versus about 72 GB for bf16. On AutoTrust's own Decision Index 0.3 public suite, the 35B System 1 scores 59.73, and NVFP4 matches 96.6 percent of the bf16 answers. That is author-run with the kit, not a board Full score.34
OpenBMB MiniCPM-V 4.7
OpenBMB's MiniCPM-V 4.7 is a 35B-A3B vision MoE, the first in that series, under Apache-2.0, and the official llama.cpp GGUF was updated this morning. Q4_K_M is 21.17 GB, a 5090 or Spark class load, not a phone pack. The iPhone 17 Pro Max 13+ tok/s claim on the card is a different runtime (LES, layer-wise quant, and DSpark). The weights have been on the Hub since October 7.56
UIBK-Embedding-v1
Abdelrahman Abdallah of DataScience-UIBK published a Hub post on October 12 for UIBK-Embedding-v1, an Apache-2.0 179M English retriever. It keeps a dense head and a late-interaction ColBERT head in one encoder, and both scores sit inside one unmodified MaxSim, so you are not maintaining two indexes. It loads as a normal PyLate ColBERT model. The author's BEIR number is 59.79 nDCG@10, and the weights have been on the Hub since October 9.78
