Nanbeige's Apache-2.0 33B teacher is fifth on Hugging Face's default trending page. Nanbeige4.3-33B is the Kanzhun / Boss Zhipin lab's flagship chat card, with a 512K context. Community llama.cpp GGUFs from bartowski and AtomicChat landed on 25 October, and Nanbeige still has no official quant. Alibaba DAMO refreshed three Apache-2.0 RynnUnify cards this morning as the paper joined today's Daily Papers board. The papers feed rotated overnight to 26 October, and RecGPT-Team's UserArena is the new shopping-agent benchmark, with GitHub and no Hub weights.
On the Hub
Nanbeige4.3-33B
Nanbeige/Nanbeige4.3-33B is fifth on the default trending page. The chat card went up on 24 October and last changed on 25 October. It is Apache-2.0, 33B parameters in BF16, 27 files totaling 66.260 GB. The companion base card, Nanbeige/Nanbeige4.3-33B-Base, is the same 24 October drop with a 131K context. The chat model is configured for 524,288 tokens.12
Nanbeige presents it as the flagship teacher for the Nanbeige4.3 series, with a smaller 3B distill still to come. At highest thinking effort the authors report Terminal-Bench 2.1 at 80.9, ahead of Qwen3.8-27B at 73.0, Muse-Glimmer-30B at 61.8, and Gemma4-31B at 47.2. They also report AIME 2026 at 95.2 and DeepSWE at 65.5. Gemma still leads GPQA Diamond at 91.3 against 89.5, and Qwen takes SWE-Bench Pro at 61.7 against 60.3. Those numbers are the authors', not an independent run.3
bartowski + AtomicChat Nanbeige GGUF
Community llama.cpp packs of Nanbeige4.3-33B landed on 25 October. There is no official Nanbeige GGUF. bartowski/Nanbeige_Nanbeige4.3-33B-GGUF went up that afternoon. The author-recommended Q4_K_M file is 20.79 GB.4
AtomicChat/Nanbeige4.3-33B-GGUF went up a few minutes earlier, with Q8_0 at 35.195 GB. These are community conversions of the 24 October teacher, not a Nanbeige-shipped quant.56
Alibaba DAMO RynnUnify
Alibaba DAMO's RynnUnify paper joined today's Daily Papers board. The paper is from 23 October, and Daily Papers dated it 26 October. Three Apache-2.0 Hub cards last changed this morning. The cards themselves date to 22 October, so this morning's timestamps are a refresh, not new weights overnight.78
RynnUnify-Base is 20 files and 31.515 GB, an any-to-any finetune of nvidia/Cosmos3-Nano at 15.75B. RynnUnify-Agentic-Reasoner is a Qwen3-VL image-text-to-text card at 17.550 GB (8.77B). RynnUnify-WAM-RoboCasa is the world-action robotics card at 32.184 GB (16.09B). The README frames it as agentic embodied intelligence for visual planning, physical reasoning, and robot control, with code on GitHub.91011
Daily Papers rotates to Oct 26
Hugging Face Daily Papers rotated overnight. Just over half the board is now dated 26 October. Tencent's WeVisOne-VL still opens the ranked list at 244 up, which is score movement on the 24 October lead, not a new drop. RecGPT-Team's UserArena is new on today's board at 34 up: a shopping-agent benchmark with GitHub and no Hub weights.12131415
ByteDance's RSIHub is also dated 26 October (24 up), a framework repo rather than a checkpoint. OmniDR-Bench dropped off the board overnight. RynnUnify is on that same 26 October board, with Hub weights.16
