Hugging Face Daily Digest Nov 1, 2026

Nanbeige published official INT4, FP8, and INT8 builds of Nanbeige4.3-33B, the dense model whose full-precision card is already at the top of the Hub. OrcaRouter posted a Mix GGUF of Mistral Large 4, and the repository reports 676.62 GiB of used storage. Bartowski added a GGUF of a Logics-MLLM software-engineering fine-tune in the Qwen3.6-35B-A3B family. Daily Papers still has no November 1 or October 31 slice, and KAIST's Look Only When Needed is the upvote lead on the October list that remains.

On the Hub

Nanbeige4.3-33B official INT4, FP8, and INT8

Nanbeige published three quantized repositories for Nanbeige4.3-33B on October 31, created between 00:45 and 00:48 UTC and last modified between 09:25 and 09:41 UTC. The INT4 card is 21.02 GiB of Apache 2.0 compressed-tensors weights.¹ The FP8 card is 33.10 GiB and the INT8 card is 33.54 GiB, in the same layout rather than GGUF.²³ The full-precision teacher is still the 61.71 GiB BF16 repository, last modified on October 28, and that card is number one on the default page.⁴

The teacher is a dense Llama of about 33.12 billion parameters, with 52 layers, a hidden size of 7168, eight key-value heads, and a 524,288-token context. The config has no mixture-of-experts field and no RoPE scaling. Only the INT4 readme describes how that quant was trained: quantization-aware training, on-policy distillation from BF16 into INT4, and reinforcement learning on the INT4 model. The FP8 and INT8 openings say only that each repository is the quantized version of the teacher.

INT4 stores 4-bit pack-quantized linears and leaves lm_head and model.embed_tokens out of that pack. FP8 is float-quantized 8-bit (F8_E4M3) and skips lm_head, and INT8 is 8-bit pack quantization with the same skip. At 21.02 GiB the INT4 file is the smallest of the three, and it is a compressed-tensors weight set rather than a GGUF sized for a laptop. Nanbeige has not posted an official GGUF with these cards.

The comparison table on all three quant readmes is the teacher's BF16 table, so those figures are not INT4 scores. On it the teacher lists 80.9 on Terminal-Bench 2.1, 23.8 on Terminal-Bench 4.0, 65.5 on DeepSWE, 60.3 on SWE-Bench Pro, 50.8 on SWE-Atlas-QnA, 89.5 on GPQA Diamond, 95.2 on AIME 2026, and 84.0 on AA-LCR. The table compares the teacher with Qwen3.8-27B, which leads SWE-Bench Pro at 61.7, Muse-Glimmer-30B, and Gemma4-31B, which leads GPQA Diamond at 91.3.

OrcaRouter Mix GGUF of Mistral Large 4

OrcaRouter published a GGUF conversion of Mistral Large 4 at 01:20 UTC on November 1, last modified at 01:56 UTC.⁵ It quantizes the existing mistralai/Mistral-Large-4-1T-A52B checkpoint under the Mistral Ecosystem Open-Weight License.⁶ Mistral's newest Large 4 repository by creation date is still the NVFP4 card from October 15, so this GGUF is a community quant of a checkpoint already on the Hub. The metadata records the architecture string mistral-large4, a context of 1,048,576 tokens, and 1,049,357,365,536 parameters.

Hugging Face reports 676.62 GiB of used storage for the repository. The file tree lists a Mix-256G folder in six shards, a Mix-512G folder in twelve shards, and an imatrix file. Those folders do not list their own byte sizes, so 676.62 GiB stays the repository total. A separate total-file-size field reads 216.94 GiB, a different measurement that the repository page does not tie to either Mix folder. Unsloth's Large 4 GGUF was last modified on October 29 and is unchanged as a release.⁷

Bartowski GGUF of Logics-SWE-Qwen3.6-35B-A3B

Bartowski published a GGUF of Logics-MLLM's Logics-SWE-Qwen3.6-35B-A3B on October 31 at 16:56 UTC.⁸ The base card was created on the morning of October 30 and last modified at 04:45 UTC that day.⁹ It is an Apache 2.0 fine-tune of Qwen/Qwen3.6-35B-A3B, 35.11 billion parameters in BF16, about 65.42 GiB.¹⁰ The base readme dates the release 2026.10.30 and points at arXiv 2609.23377.

This GGUF is bartowski's quant of the Logics-MLLM software-engineering fine-tune, in the Qwen3.6-35B-A3B family. It is not a new card from the Qwen organization. The newest model card there is still Qwen-Image-2.1-Turbo, from October 9.¹¹

Daily Papers

Daily Papers has no November 1 slice and no October 31 slice. Both date pages are empty, and the current list of 50 is still 39 papers dated October 30 and 11 dated October 29.¹² On that October 30 set, KAIST AI's Look Only When Needed (2610.36671) is the numerical upvote lead at 100, up from 94, with 13 GitHub stars, up from 10.¹³

inclusionAI's Intrinsic Distribution Matching (2610.39633) now shows 20 upvotes, where the same count was 239 on October 31, while GitHub stars on inclusionAI/IDM moved from 59 to 112.¹⁴¹⁵ The paper is still on the list. The upvote change is not described there as a withdrawal or a retraction. Xiaomi's A-RSI paper (2610.38568) moved from 158 upvotes to 71 on that same October 30 list.¹⁶