Three releases rotate through Hugging Face trending rank 30

From Oct. 4 to 9, three different models briefly held Hugging Face trending rank 30, including a 7.89 GB Saluki Qwen GGUF.

Hugging Face trending rank 30 rotated among three unrelated releases from Oct. 4 to Oct. 9, 2026. On Oct. 4, nerkyor published a multi-round SFT and RLOO post-train of official Qwen3.8-27B reporting static FP8 scores of GPQA 178/198, MMLU 450/500, and LiveCodeBench 90/100 versus 177/444/83 for the original FP8 checkpoint, with 94K truncations falling from 4 to 1 on GPQA and from 13 to 3 on LCB, and DFlash2 decoding at about 180 tokens/s versus 46 without. On Oct. 8, Conway Research released Underdog-Saluki-27B-1.0, a 7.89 GB IQ2-mix GGUF of Qwen3.8-27B for stock llama.cpp under Apache 2.0; it scored 88 versus 84 for the full 54 GB model on a 120-task BFCL v4 subset and 42 versus 35 on 100 parallel tool-call tasks, with about 96% average retention across nine benchmarks. MarkTechPost's Oct. 9 profile agrees the 7.89 GB file is a 2-bit GGUF that beats the roughly 54 GB original on tool calling and scores lower on competition math and reasoning, but it does not repeat those bench counts; the model card separately cites competition-math retention of roughly 82 to 85 percent and an AIME 2025 comparison of 96.7 versus 79.2 from a different harness. The same day, Qwen-Image-2.1-Turbo, an eight-step accelerated 7B checkpoint of Qwen-Image-2.1 that defaults to classifier-free guidance of 1 and prefix KV caching through QwenImage21Pipeline, was also reported trending near number 30.

• On 2026-10-04, nerkyor's Qwen3.8-27B SFT/RLOO post-train reported static FP8 scores of GPQA 178/198, MMLU 450/500, and LiveCodeBench 90/100 versus 177/444/83, with 94K truncations falling from 4 to 1 on GPQA and from 13 to 3 on LCB.
• That repo ships BF16, FP8, NVFP4, INT8, INT4, and GGUF builds; MTP and DFlash2 speculative decoding were about 91 and 180 tokens/s versus 46 without.
• Underdog-Saluki-27B-1.0 is a 7.89 GB IQ2-mix Apache 2.0 GGUF of Qwen3.8-27B for stock llama.cpp; MarkTechPost calls the same file a 2-bit GGUF against an original of about 54 GB.
• On a 120-task Underdog Bench subset of BFCL v4, Saluki scored 88 versus 84 for the full model, and 42 versus 35 on 100 parallel tool-call tasks, with about 96% average retention across nine benchmarks.
• Competition-math retention is roughly 82-85%; a public AIME 2025 figure cited for the full model is 96.7 versus 79.2 for Saluki, from a different harness. Optional vision needs a separate 629-928 MB mmproj file.
• On 2026-10-09, Qwen-Image-2.1-Turbo, an eight-step accelerated 7B text-to-image and editing checkpoint, was trending near #30, defaulting to CFG 1 and prefix KV caching via QwenImage21Pipeline.