Hugging Face Daily Digest Nov 2, 2026

CCORE replaced the quant files on a new Qwen3.5-4B card aimed at GLB scenes, and Nota AI posted a gated prune of Mistral Large 4 that removes 20 percent of the routed experts. Daily Papers opened a November 2 list of 22 papers, led by a Tsinghua note on what a light System-1 filter lets a heavy System-2 do. ServiceNow's Apriel Voice 27B paper sits on that same list, while the weights on the model card remain the October 29 upload.

On the Hub

CCORE Qwen3.5-4B GLB 3D coder

CCORE published Qwen3.5-4B-GLB-3D-Coder at 22:30 UTC on November 1, then replaced the Q4_K_M and Q8_0 files and added a BF16 vision projector at 04:50 and 04:52 UTC on November 2.¹ The card is Apache 2.0, tagged for text generation, with the architecture string qwen35 and a context of 262,144 tokens. It is a community fine-tune of Qwen3.5-4B, and the readme calls it an independent CCORE model, not an official Qwen release.² The newest model card under the Qwen organization is still Qwen-Image-2.1-Turbo, from October 9.³

The repository holds Qwen3.5-4B-GLB-3D-Coder-Q4_K_M.gguf, Qwen3.5-4B-GLB-3D-Coder-Q8_0.gguf, and mmproj-Qwen3.5-4B-BF16.gguf. Hugging Face reports 14.03 GiB of used storage for the whole card, while a separate 2.52 GiB figure describes one GGUF rather than the repository. The readme says those filenames stayed the same when the quants were replaced, that full Safetensors weights are not in this repo yet, and that training is still running. The stated aim is Three.js scenes and assets, with GLB export for editors such as Godot, and optional image inputs through the BF16 projector.

The older sibling, Qwen3.5-4B-3D-Coder, is number 7 on the default page.⁴ Its newest update today only adds community support links to the readme, that parent already held the same quant names, and the GLB card is not on that page.

Nota AI prune of Mistral Large 4

Nota AI published Mistral-Large-4-1T-A52B-NVFP4-Global-Pruned-20 at 05:33 UTC on November 2, last modified at 05:39 UTC.⁵ The readme describes Nota's non-uniform expert pruning, which removes 20 percent of the routed experts from Mistral's October 15 NVFP4 checkpoint.⁶ It lists that original NVFP4 at 568 GiB and the pruned checkpoint at 460 GiB, and it says the result holds a 512K context on four B200s with tensor parallelism of 4. Hugging Face measures the repository at 459.96 GiB of used storage. The 568 GiB and 460 GiB numbers stay readme claims, and 459.96 GiB is the size of the repo.

Readers have to request access through a notification form before they can pull the weights, the license card reads other, and the card is tagged for vLLM, NVFP4, mixture of experts, and expert pruning. The file tree includes safetensor shards from consolidated-00001-of-00032 through 00032, and those names plus the 459.96 GiB storage figure are what show the weights are on the Hub.

Mistral's newest Large 4 card by creation date is still the October 15 NVFP4.⁷ Unsloth's Large 4 GGUF was last modified on October 29, and OrcaRouter's Mix GGUF was last modified at 01:56 UTC on November 1, so this prune is a separate compression of the NVFP4 weights.⁸⁹

Daily Papers

Daily Papers has a November 2 date with 22 papers, each marked as submitted to the daily page at midnight UTC on November 2.¹⁰ The November 1 and October 31 date pages are still empty.¹¹¹² An unfiltered list of 50 is now 22 papers dated November 2 and 28 dated October 30. ArXiv dates on the new 22 run from October 22 through October 30, so the slice is newly featured rather than written overnight.

The upvote lead is Tsinghua's "Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?" (2610.39497), at 52 upvotes, with an arXiv date of October 29.¹³ The paper asks whether the next gain comes from co-designing a heavy System-2 with a light System-1 filter, rather than from scaling System-2 alone. The Hub page lists no models and no GitHub repository. A project page is up, and no new Tsinghua checkpoint appeared with the feature.¹⁴

R²Orch (2610.22740), from CUHKSZ, shows 39 upvotes.¹⁵ Tencent's "Wizard: Manipulate Any World in Real Time via Agentic Control" (2610.40339) shows 22 upvotes, with a project page and no matching Tencent checkpoint.¹⁶¹⁷ Apriel Voice is also on this slice, and the model card is the next brief.

Apriel Voice 27B

ServiceNow-AI's Apriel Voice 27B preview is on the November 2 papers list as 2610.40738, with 14 upvotes and an arXiv date of October 29.¹⁸ The model repository was created on October 13, and updates at 05:21 and 05:27 UTC on November 2 changed the readme and a bibtex file, while the weights are still the upload from 16:55 UTC on October 29.¹⁹

The card is MIT and not gated. It takes audio plus text and returns text, the weights are BF16 at about 27.23 billion parameters, and the repository uses 50.75 GiB. The readme describes speech understanding, text responses, and tool calls through a shared Gemma 4 backbone, while text and image inputs are still accepted and speech out is left to an external synthesizer.