The Hub is quiet on new base weights today. The news is a same-byte quantization bake-off, a sparse-attention gate drop, and two Daily Papers. ARCHsloth posted a FINAL-Bench AutoRound comparison with two corrected flags. Tencent Hunyuan put Simple-Attention-Sparsification gates on the Hub. Daily Papers lead with Benchmark Radar from CMU and LIT from NUS.
On the Hub
ARCHsloth: AutoRound two-line fix, same-byte Q4_K_M
FINAL-Bench published a community blog on September 14, 2026. It compares AutoRound quantizations against Unsloth Q4_K_M at identical byte counts1. Two flags were wrong for months. The correct scheme is GGUF:Q4_K_M, not W4A16. The fix also enables --enable_alg_ext.
That enables SignRoundV2 rounding search. The blog claims KL drops of 20 to 54 percent on Qwen3-4B versus Unsloth Q4_K_M. Lower KL is better. The numbers are Korean -54.4 percent, source code -52.9 percent, Japanese -33.7 percent, English -33.2 percent, Chinese -31.1 percent, Spanish -30.2 percent, Russian -25.0 percent, Arabic -23.2 percent, Thai -21.5 percent, Hindi -20.9 percent.
Hub files landed today under Archsloth. The 4B collection is at https://huggingface.co/Archsloth/Qwen3-4B-GGUF2. The Q4_K_M file is 2,497,280,800 bytes. It is byte-identical to Unsloth. The 9B collection is at https://huggingface.co/Archsloth/Qwen3.5-9B-GGUF3.
The 27B collection is at https://huggingface.co/Archsloth/Qwen3.8-27B-GGUF4. The 27B Q4_K_M was not shipped. The blog says it lost English by 15.8 percent versus Unsloth UD-Q4_K_M. The shipped 27B rung is Q6_K.
Tencent Hunyuan SAS: router-only AttnGates for frozen Qwen3
Tencent Hunyuan created https://huggingface.co/tencent/Simple-Attention-Sparsification on 2026-09-14T07:43:25Z5. It is router-only AttnGates for frozen Qwen3-4B, 8B and 14B. The paper is 2609.13141. The HF Papers page is https://huggingface.co/papers/2609.131416. The repo contains gate weights, not a standalone model.
Each size ships attn_gate_weights.pth plus config and tokenizer. The backbone stays frozen. Training uses a KV block of 64 tokens. A sparse decode budget of 2,048 tokens is used. Inference requires the sglang-blocksparse fork with seer_attn backend.
The Hub card does not state a license. No license is listed on the model page.
Benchmark Radar: CMU living catalog of evals
Benchmark Radar is #2 on HF Daily Papers today. The paper is arXiv 2609.11115. The HF Papers page is https://huggingface.co/papers/2609.111157. It is a living catalog of AI benchmarks from Carnegie Mellon University. The paper claims 1,283 source records.
It claims 12,916 score observations across 790 records. Daily discovery is from 37 sources. The project site is benchmark-radar.org. The catalog covers LLM eval, agent tool use, coding, reasoning, safety and domain specific work.
LIT: NUS latent interface for robot foundation models
LIT is #1 on HF Daily Papers today. The paper is arXiv 2609.12641. The HF Papers page is https://huggingface.co/papers/2609.126418. It is Latent Interface Training from National University of Singapore. Stage one trains an action expert on language, state and terminal SE(3) pose with no images.
Stage two adds a pose-supervised latent interface as the only visual path. The paper claims LIBERO-Plus gains of +3.87 to +10.70 percentage points. Real-world tests claim +13.30 to +16.70 pp under unseen cameras and lighting. No Hub weights are linked on the paper page.
