Hugging Face Daily Digest Oct 30, 2026

Mistral Large 4, nicknamed Le Chonk, is the weight drop on the Hub this morning. The FP8 card is on the default page, Pika published a new video adapter beside it, and Hugging Face opened a science-data initiative called Project ProtoStar. Daily Papers turned over to October 30. A UIUC benchmark paper and inclusionAI's Intrinsic Distribution Matching lead that slice, and neither paper links a model.

On the Hub

Mistral Large 4 (Le Chonk)

Hugging Face published the open weights for Mistral Large 4 on October 29, under the nickname Le Chonk.¹ The FP8 repository mistralai/Mistral-Large-4-1T-A52B entered the default trending page at number 11.² It is a multimodal mixture of experts that takes text and images and returns text, with a one-million-token context. The Hugging Face post calls the model about 1 trillion parameters in total, with 49 billion active, the repository id uses an A52B label, and Mistral's current docs card lists 52 billion active parameters and 1.05 trillion total.³

The public API arrived earlier in October, on a promise that the weights would follow by the end of the month. These Hub cards are that weight release. The post names two official variants: the FP8 card now on the default page, and an NVFP4 sibling that uses the FP8 repository as its base.⁴ In the post's words, the license is the Mistral Ecosystem Open-Weight License, an MIT derivative that limits model-as-a-service offerings. A specialized cybersecurity build that Mistral offers to design partners under contract is a separate preview, and it is not this checkpoint.

The same post, using figures as of October 29, places Large 4 fifth among open-weight models on the Artificial Analysis Intelligence index. Its coding figures are 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, 28.3 percent on Terminal-Bench 4, and a combined Coding Agent Index of 49.8 percent. Knowledge-work figures on that page are 59.9 percent on AutomationBench and 1,393 Elo on AA-Briefcase. On cyber benchmarks it puts the model in the top five of the Artificial Analysis Cyber Index, at 82 percent on the reproduce-then-patch item, 93 percent of 40 Cybench challenges, and 93.3 percent attack resistance on Lakera B3. The post also says the average refusal rate on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than for any other open-weight model.

For local and hosted runs, the post points at Transformers, vLLM, llama.cpp, TRL, Inference Endpoints, and Dell Enterprise Hub, and the llama.cpp example serves the ggml-org Q2_K quant. Unsloth published a GGUF repository on October 29, and the smallest summed folder there, UD-IQ1_S, is 217.01 GiB across six shards.⁵ The Q2_K build the post points at, on ggml-org, is 354.56 GiB.⁶ A 2.09 GiB file at the root of the Unsloth repo, dspark-Mistral-Large-4-1T-A52B-Q8_0.gguf, is a sidecar rather than the full model, and none of these builds fit a 128 GB machine.

Pika-Labs Pika-H3-Flex

Pika-Labs published Pika-H3-Flex on October 29, and the card is number 29 on the default page.⁷ It is an adapter on MiniMaxAI/MiniMax-H3, licensed under the MiniMax H3 community agreement, and it is the lab's first public model on the Hub. The README describes a post-trained H3 model that generates video with synchronized audio in six sampling steps and can take up to twelve reference inputs. The repository also holds Comfy LoRAs plus fl2va and ref2va adapters.

Open-weight inference on the Hub repository is native 480p and 768p. Higher resolutions, including 1080p, 2K, and 4K, are what Pika claims for its hosted API, not for the local weights. The card has recorded no downloads, so a local runtime has not shown up yet.

Project ProtoStar

Hugging Face announced Project ProtoStar on October 29.⁸ The post describes an effort to put publicly funded datasets, models, and code into formats that models can use. Founding university partners named there are Carnegie Mellon, Purdue, Georgia Tech, Arizona State, and George Washington, through the Open Forum for AI, along with Astera, Hugging Face, and ORCA. The first challenge, hosted by Carnegie Mellon, opens on November 2. The post does not publish a model id, and the challenge is not open yet.

The post gives its own scale. It puts U.S. federal research and development at about $190 billion a year, and it counts Hugging Face among ten NIH generalist repositories. It says the Hub holds more than 17,000 public science datasets, or about 29,000 if medicine and mathematics are included. The same account says the Hub added more than three times as many new science datasets in the first nine months of 2026 as in the same months of 2025.

Daily Papers for October 30

Of the latest fifty Daily Papers, twenty-three are dated October 30. UIUC's paper, Your Benchmark Is an Empirical Measure Over Difficulty, leads that slice with 137 upvotes and no linked model.⁹ It treats a benchmark score as a measure of difficulty and names mismatch and coverage gaps as ways that score can mislead.

inclusionAI's Intrinsic Distribution Matching follows at 133 upvotes, again with no linked model.¹⁰ It adds moment constraints to distribution-matching distillation for few-step text-to-image. KAIST's Look Only When Needed is next on the same date, at 56 upvotes, and it also has no linked model.¹¹ The top of the October 30 slice is papers, not weight drops.