Hugging Face

Jina AI 发布文档解析模型 jina-ocr-v1，可以把 PDF、扫描件、表格和图表直接转成 Markdown。它基于 DeepSeek-OCR 做后训练，沿用其约 34 亿总参数、解码时每个 Token 激活约 5.7 亿参数的 MoE 架构，并加入 FastMTP 推测解码。 Jina 自测中，jina-ocr-v1 在 OmniDocBench v1.6 得 91.14，比 DeepSeek-OCR-2 高 0.89 分；在 olmOCR-Bench 得 83.4，比 DeepSeek-OCR 高 7.4 分。 吞吐量是 Jina 主打的卖点之一。单张 A100、并发 32 时，它达到 2.57 页/秒，在 Jina 测试的 14 个系统中最高，比 DeepSeek-OCR 的 2.10 页/秒高约 22%。 模型权重已经放上 Hugging Face，采用 CC BY-NC 4.0，商业使用需要联系 Jina。

@JinaAI_

Announcing jina-ocr-v1, our new visual document parser with 3.4B total parameters and 570M active parameters, with speculative decoding built in. Throw PDFs, scans, tables, charts, or invoices at it and get clean markdown back. Available on 🤗 & Jina Reader `x-respond-with` today

this is so useful! if multiple diffusers models share the same text encoder or vae, now they will share the same blobs in your cache

@victormustar

Hugging Face cache is way more than a folder full of downloaded models 🤫 > With huggingface_hub v1.32, identical Xet-backed files across repos are stored only once, so if 5 repos share the same 20 GB weights file: > -> before: ~100 GB on disk -> after: ~20 GB > And once the blob is local, another repo can reuse it with zero payload download 💃 > Xet already made the remote storage content-addressed. Now that identity extends into your local cache too.

New open weights from @yandexdotcom: AliceAI-Foundation-80B-A3B-Base, now on Hugging Face. • Trained from scratch • 80B total / 3B active parameters • 262K-token context • Apache 2.0 open weights A pretrained base model for fine-tuning.

• huggingface.co

the bitter lesson is that most AI researcher spin-outs are basically just acquihire opportunities and their products and research are almost all completely worthless

@harshagundal

They were building in stealth for 2 years, I was building in stealth for 2 hours… > Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe. > ⚡️Demo below on a M4 MacBook⚡️ > every LLM has the ability to efficiently batch inference every key of a JSON at the same time and generate probabilities from a set of possible categories. No new training required, but it’s easy to optimize if you need! > On hugging face now!

Hugging Face 已封禁 AI 安全公司 Audn 上传的 penclaw-GLM-5.3-abliterated-for-offensive-cyber，页面显示其违反 Content Policy。这个版本直接修改 GLM-5.3 权重，削弱模型的拒答行为。原仓库名和介绍还明确写着 offensive cyber。作者随后删掉这几个字，重新上传了模型。 具体为什么被封，Hugging Face 没有公开解释。作者称自己也没搞懂原因，社区有人猜测是 offensive cyber 的命名触发了审核。Hugging Face 的现行政策确实禁止旨在破坏、未经授权访问系统，以及生成恶意代码的内容。 但这类削弱模型拒答限制的版本在 Hugging Face 并不少见，平台上已有数千个类似的 abliterated 模型。

@audn_ai

Hello world! It was probably taken down because of its name. Similar content was reuploaded here: > No, we will not do PR by saying it was so "harmful" that it was taken down. > Timing is interesting because we were also banned by @OpenAI recently for trying to use their model on OWASP Juice Shop ( A vulnerable GitHub web application for red-teaming training ) . We were also never accepted for Anthropic's program, even though we applied. > We applied several times for both companies' Trusted Access for Cyber program and were rejected

• huggingface.co

Introducing: a coding agent (Pi) running entirely in your browser using a 2B model on WebGPU 🤯 MiniCPM5-2B + Pi, powered by Transformers.js + WebGPU + 4-bit ONNX weights. All previous attempt to create this failed but MiniCPM5 seems to make it usable. Available now on Hugging Face 👇

Xiaomi-CocktailASR-1 just landed on @huggingface, many people yapping, one clean transcript 🍸 feed an audio with people speaking over each other, a clip of the voice to isolate, get the transcription from just that person ▶️ on Spaces

• hf.co

THE SANDBOX WAS A PROP! OpenAI Turned Off the Guardrails, Left a Door to the Internet, and Then Sold the Hugging Face Breach as “Rogue AI” It is time to understand how you were lied to and by whom. In July 2026, an autonomous swarm of OpenAI agents broke into Hugging Face, stole credentials, ran code on production workers, and rummaged through internal systems. The official story was that the models “went rogue.” The paperwork says something colder. The labs asked for this. Now the story can be told. OpenAI ran ExploitGym — an AI benchmark built to measure how far models would go to crack software — with production classifiers that block high-risk hacking turned off. Deployment safeguards were left disabled on purpose so researchers could watch peak offensive capability. GPT-5.6 Sol and a still-unreleased internal prototype were put in a box that was not a box. They were allowed to talk to an internal package-cache proxy. This is not a real world test, it is a setup with predictable outcomes. That pro

zoom 256× into a photo, one 4× step at a time 🔍 OracleZoom approaches super-resolution by writing its own prompt at every level and re-rendering ▶️

• hf.co

I trained an AI model on my phone through Telegram > Using OpenClaw running on Hugging Face infra: ML Claw > It beat DeepSeek V4 Pro on the given task while having 80 thousand times less parameters > That's right. The model, GoePT-1-20m is only 20 million parameters. It runs in the browser, on the CPU. And it beats DeepSeek V4 Pro, a 1.6 trillion parameter model, in AlmanBench > This concludes my 4 year old side project (fun fact, I created a dataset for this pre AI agents, using SpaCy, and paid 50 euros from my own pocket to rent 4090s on runpod to train a T5 variant, *years* before I joined Hugging Face. That first attempt was not very successful. This one is. My first ML adventure 🤗) > The goal of this side project was to show that you can *vibe* machine learning now, on platforms with tightly integrated GPUs, storage and compute, like Hugging Face > Including dataset creation, autoresearch and the final training run (hat tip to ML-Intern which g

@onusoz

I trained an AI model on my phone through Telegram > Using OpenClaw running on Hugging Face infra: ML Claw > It beat DeepSeek V4 Pro on the given task while having 80 thousand times less parameters > That's right. The model, GoePT-1-20m is only 20 million parameters. It runs in the browser, on the CPU. And it beats DeepSeek V4 Pro, a 1.6 trillion parameter model, in AlmanBench > This concludes my 4 year old side project (fun fact, I created a dataset for this pre AI agents, using SpaCy, and paid 50 euros from my own pocket to rent 4090s on runpod to train a T5 variant, *years* before I joined Hugging Face. That first attempt was not very successful. This one is. My first ML adventure 🤗) > The goal of this side project was to show that you can *vibe* machine learning now, on platforms with tightly integrated GPUs, storage and compute, like Hugging Face > Including dataset creation, autoresearch and the final training run (hat tip to ML-Intern which g

This is a massive leap for open-source AI! 🚀 A complete fleet ranging from 0.9B to 375B parameters with top-tier frontier performance, plus full transparency on training data and code. Game changer! 💡🔥

@IFM_AI

Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters. > - Frontier performance: Across coding and agentic tasks, K2 Horizon delivers top-tier performance in every size class—with the 0.9B, 3.7B and 7B models setting new state of the art at their respective scales. - Radical openness: K2 Horizon represents the largest fully open-source model launch in AI history. The fully open code, training data and recipes are a significant step forward in transparency. > Launch page: Tech blog: Hugging Face:

• ifm.ai
• ifm.ai
• huggingface.co

🚨 Episode 94 of The Intelligence Journal is here. NASA Open-Sourced an AI Model of the Moon 🌙🛰️ NASA and IBM just released a foundation model trained on the Moon itself. Two million image tiles. Seventeen years of Lunar Reconnaissance Orbiter data. Over a million high-resolution frames at one metre per pixel, plus nearly 964,000 multispectral images, stitched together with data from GRAIL, Lunar Prospector and Japan's SELENE. It maps craters, dates surfaces, spots volcanic features that break existing cooling models, detects changes between passes, and estimates where ice is stable near the poles. It beat the baselines on ice prospecting 🧊 And they put the whole thing on Hugging Face with the code on GitHub. Free. Join us as we break down: - Why a general model trained once on raw sensor data beats a decade of purpose-built algorithms, and why that lesson transfers to every domain sitting on unused data. - What finding stable ice actually unlocks, because water on the Moon is fuel, air and the differe