Hugging Face Daily Digest Oct 31, 2026

Diffuspace wrote up Clover-35B-A3B-Preview on October 30, a diffusion language model whose weights have been on the Hub since October 24. The same day, NVIDIA published an NVFP4 quantization of DeepSeek-V4.1-Flash-1017, and those files add up to 399.55 GiB. Tencent's Youtu-Parsing-Omni paper moved into the top of Daily Papers beside a checkpoint that has been on the Hub since October 9. The daily list still has no October 31 slice: inclusionAI's distillation paper now links a community demo, and Xiaomi's A-RSI climbed without a model.

On the Hub

Clover-35B-A3B-Preview

Diffuspace published the community post for Clover-35B-A3B-Preview on October 30.¹ The weight card Diffuspace/Clover-35B-A3B-Preview has been public since October 24, and its last Hub update is still October 28, so the post is what arrived, not a first upload of the weights.² The post says the lab is officially releasing the model, while the card date is six days earlier. Both the weight card and the GGUF card use Apache 2.0, and the base family named on the card is Qwen3.6-35B-A3B.

Clover is a mixture-of-experts diffusion language model, about 35 billion parameters in total and about 3 billion active, with a thinking mode. The config is custom clover_moe code, with 256 experts and 8 of them used on each token, and the context length is 131,072 tokens. A safetensors count on the card is about 34.66 billion BF16 parameters, which is the total behind the 35 billion figure, not the active count.

The benchmark numbers are the post's own, and they have not been rerun. It prints 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, 65.3 on τ³ Overall, 83.8 on GPQA Diamond, 95.8 on ZebraLogic, and 72.0 on AA-LCR. On speed, the post reports 680 tokens per second on one NVIDIA GB300 in BF16 at batch size 1, and 515 tokens per second against 365 for a similarly sized Qwen setup with DFlash2 on one H200. The post names SGLang and llama.cpp for inference, and it says a decision-model variant will come later.

The official GGUF repository, Diffuspace/Clover-35B-A3B-Preview-GGUF, was created on October 29 and lists a 128K context.³ Q4_0 is 18.24 GiB, Q4_K_M is 21.42 GiB, and Q8_0 is 35.26 GiB. The post also says an INT4 build runs at 101.8 tokens per second on MATH-500 on a 16 GB 5070 Ti, against 48.5 in the comparison it prints. That INT4 file is not in this GGUF tree.

DeepSeek-V4.1-Flash-1017 in NVFP4

NVIDIA published nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 on October 30.⁴ It is a Model Optimizer quantization of deepseek-ai/DeepSeek-V4.1-Flash-1017, which remains the top model on the default Hub list.⁵ NVIDIA does not own the underlying model. The license is MIT, the README says the quant is ready for commercial or non-commercial use, and the card shows no downloads.

The quant fields do not use one label. Expert weights are NVFP4 (moe_quant_algo), the algorithm is mixed precision, and the quant_method string in the config is still fp8, while the card tags say FP4. Experts are 4-bit with a group size of 16, and attention, shared experts, the output head, and the MTP weights are left out of that quant. The tree sums to 399.55 GiB across 48 shards, which does not fit a 128 GB machine.

The README describes the base as a multimodal mixture of experts with a causal encoder-decoder, CSA2, and a context of up to one million tokens. This file is the 1017 checkpoint, not NVIDIA's earlier NVFP4 of DeepSeek-V4.1-Flash and not the FP4 build of Qwen3.8-Flash-Next. It is also separate from the ISTA-DASLab GSQ-RCO GGUF and from Unsloth's GGUF builds of the same 1017 base.

Youtu-Parsing-Omni

Tencent's paper, Youtu-Parsing-Omni: One Encoder, One Schema, All Modalities, entered the top of Daily Papers with 124 upvotes.⁶ The summary on the paper card calls it a unified parser for documents, images, audio, and video that shares one set of weights. The paper's publish date is October 29, and it was submitted to Daily Papers on October 30.

The linked checkpoint, tencent/Youtu-Parsing-Omni, has been on the Hub since October 9, and the card was touched again on October 30.⁷ The license field reads other, and the paper links that one model. What changed is the paper's place on the daily list, not a new set of weights.

Daily Papers

Daily Papers has not opened an October 31 slice. A request for that date returns an empty list, and of the latest fifty papers, 39 are dated October 30 and 11 are dated October 29.⁸

inclusionAI's Intrinsic Distribution Matching leads that list, at 239 upvotes, up from 133.⁹ It adds moment constraints to distribution-matching distillation for few-step text-to-image, and the public code repository lists 59 stars.¹⁰ The paper now links akhaliq/MOSS-Image-IDM-2-Steps, a community demo created on October 30 under Apache 2.0, based on OpenMOSS-Team/MOSS-Image-SFT and a four-step LoRA adapter.¹¹ That demo is not an inclusionAI checkpoint.

Xiaomi Research's A-RSI, Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement, is at 158 upvotes, up from 39, and it still has no linked model.¹² The summary says the method refreshes training tasks when the learning signal drops and uses program-generated answers, so the loop does not need external labels. UIUC's paper, Your Benchmark Is an Empirical Measure Over Difficulty, sits at 143 upvotes with no linked model, and it is no longer the lead.¹³