The Hub is quiet this morning. Four items set the day. ConwayResearch has a 27B Qwen3.8 GGUF that fits in 8 GB. FINAL-Bench opened a GPU-only quantum challenge. A Gemma 4 LoRA reads GUI click coordinates as slots. HAL-X AI announced a 322M decision model.
On the Hub
Underdog Saluki 27B 1.0
ConwayResearch put Underdog-Saluki-27B-1.0 on the Hub on October 8 and updated it on October 10. It is an Apache-2.0 IQ2-mix GGUF of Qwen3.8-27B for stock llama.cpp. The file is 7.89 GB, versus about 54 GB for the full-size weights. On the author's 120-task BFCL v4 tool-calling set it scores 88, against 84 for the full-size model. AIME keeps about 82 to 85 percent of the public full-size scores.12
FINAL-Bench Open Quantum Challenge
FINAL-Bench published The Open Quantum Challenge on October 11, with a Hub Space the same day under an MIT license. Two seasons run in parallel: GPU quantum simulation scored against a defined ground truth, and a quantum error-correction decoder that infers errors from syndromes. The prize pool is $2,000. No quantum hardware is required: a GPU and the public harness are enough. The Space includes a participation guide for a coding agent.34
GUI-Decisions-31B
Logesh Kumar Umapathi put GUI-Decisions-31B on the Hub on October 10. It is an Apache-2.0 LoRA on Gemma 4 31B. The method reads click coordinates from slot distributions instead of decoding JSON token by token. The author claims more than three times faster grounding than baselines that skip reasoning, and 25 times versus the same model with reasoning. A 2B sibling and a live Space are on the Hub.56
THX-01
HAL-X AI published a community blog on October 10 announcing THX-01, a 322M Apache-2.0 non-autoregressive decision model. The weights have been on the Hub since September 23. Six typed jobs take labels at inference time: choice, yes/no, score, number extract, excerpt, and citation. The author reports 98.4 percent on an internal 2,843-ticket set, against 97.4 percent for TypeSafe Jev 1.13, with about 10 ms on a local GPU.78
