The Hub is quiet on new weights. The news is three write-ups. One adds a decision-model endpoint to llama.cpp. One opens a scientific report generator that has been on the Hub since February. One describes a Pi coding fine-tune from last week.
On the Hub
llama.cpp now serves decision models
ggml-org posted on October 2 about a new server endpoint for decision models. The server now exposes /v1/systemone. A model can return option probabilities in one forward pass without generating a paragraph to parse.1
The post describes five GGUF decision models bundled in a collection. The GGUF cards were created on October 1 evening UTC. The collection is Apache 2.0 except OpenJev which is CC BY-NC 4.0.
The smallest model is 144M and the largest is 27B. A 5090 or a DGX Spark can run them locally. The blog notes claimed latency on an RTX PRO 6000 from 3 ms for Julia-1 to 43 ms for OpenJev.
AstaBrief 8B open scientific reports
AllenAI posted on October 2 about AstaBrief. It is an 8B model for fast scientific report generation. It starts from Qwen3-8B.2
Training used supervised fine-tuning then direct preference optimization. The Hub card allenai/AstaBrief_8B was created on February 9. The card was last modified on October 2. The license is Apache 2.0.
The blog claims fast mode produces a report in about 51 seconds versus about 178 seconds for thinking mode. An 8B model fits easily on a 5090 with 4-bit quant.
Qwen3.8-27B-pi effort-ordered coding
bytkim posted on September 30 about Qwen3.8-27B-pi. It is a fine-tune for the Pi terminal coding agent. The model is effort-ordered so lower effort levels do not spend more reasoning tokens than higher levels on the same task.3
The Hub card bytkim/Qwen3.8-27B-pi was created on September 29. The license is Apache 2.0. The author claims Terminal-Bench 2.1 Pi medium matches base xhigh at 67 out of 89 with about 41 percent fewer output tokens.
A 27B model runs on a 5090 with 4-bit quant and fits a DGX Spark. The blog is the update, not a new release.
