We're opening a Hugging Face office in Tokyo! Our goal: help open-source AI develop in Japan and grow the local community. Let's meet! ハギングフェイスの東京オフィスがオープンしました！ 私たちの目標は、日本におけるオープンソースAIの発展を支援し、ローカルコミュニティを育てることです。ぜひお会いしましょう！

Transformers v5.4.0 is live with 13 new models! Key highlights: • VidEoMT: 160 FPS video segmentation • Document AI suite: UVDoc, SLANeXt, PP-OCRv5 • PI0: Vision-language-action model for robotics • Mistral 4: Unified instruct/reasoning model • Jina Embeddings v3: Multi-task text embeddings • CHMv2: Forest canopy height mapping ~10 new model architectures contributed by the PaddlePaddle team 🤯 Plus major performance improvements and Flash Attention 4 support!

perhaps time to switch from antigravity to codex for my course

PyTorch Conference Europe is happening next month, and many of us from the Hugging Face team will be there We'll be talking about Kernels, Transformers, Audio models, processors, Diffusers, Transformers.js among others We'll be sharing a few sneak peaks of what's coming next 🤫

We just brought our community work into one place, making it easier to find events and programs like Codex Ambassadors and Codex for OSS. There are 15 community-led events coming up around the world, with more to come! Where would you like to see a Codex meetup next? 🌏

Subagents are GA in the Codex App and CLI, it's positively changed the way I work - go give it a shot now!!

Defuddle now returns Youtube transcripts! Paste a YouTube link into defuddle.md to get a markdown transcript with timestamps, chapters, and pretty good diarization! ...or if you just want to read it, try the new Reader mode in Obsidian Web Clipper powered by Defuddle.

On March 11, at the OpenAI Codex Community Meetup – Seoul, I edited and uploaded the presentation video “Effective Harnesses for Long-Running Codex: The Era of AI Coding Through the Night” by speaker Goobong Jeong. Since this is a video I edited myself, there may be some rough edges, but I hope it helps those who were curious about the talk. If you're interested in the #Ralphthon recap or learning practical approaches for maintaining long-running Codex sessions with effective harnesses, I highly recommend checking out the video below. ▶ Video: youtube.com/watch?v=tjEVBcPT… English subtitles (CC) are available in the video. @OpenAIDevs @OpenAI

On March 11, at the OpenAI Codex Community Meetup – Seoul, I edited and uploaded the presentation video “OMX — Hacking Codex (Demo Dev Story)” by speaker Yeachan Heo. (@bellman_ych ) Since this is a video I edited myself, there may be some rough edges, but I hope it helps those who were curious about the talk. If you're interested in how Codex was extended and customized through OmX, or if you're building your own agent harness, this talk is definitely worth checking out. ▶ Video: youtube.com/watch?v=XwFoegzi… English subtitles (CC) are available in the video. @OpenAIDevs @OpenAI

> Seoul’s three Codex Ambassadors (@swhan0329, @w00jun_, Junho Kong) — for organizing two (!) community meetups, where builders traded notes on everything from agent loops to OSS Codex extensions to CI workflows, and for helping connect us deeply with the local community.

hi mum, i’m on a livestream

Excited to introduce Codex for Open Source! 🔥 TL;DR - ChatGPT Pro, Codex, and API credits for eligible open-source maintainers Open source has shaped modern software, and so much of it depends on maintainers doing steady, often invisible work to keep critical projects healthy. That is a big part of why expanding our support for open source through Codex feels so natural. At OpenAI, supporting open ecosystems has long been part of our story, from Gym and Universe to Triton, Whisper, and CLIP, and more recently gpt-oss open-weight models. Last year, we launched the Codex Open Source Fund to support projects with API credits. Today, we’re building on that by giving more maintainers access to ChatGPT Pro, Codex & Codex Security. This is, at its core, a thank you ♥️ Maintainers do extraordinary work for the broader software ecosystem. If Codex can help reduce some of that load, whether through reviews, debugging, tests, or understanding large codebases, this feels like a meaningful way to give back. Go apply today! 👇

how to use obsidian claude code to build a 24/7 personal operating system and build your startup: 1. write everything in markdown (daily notes, projects, beliefs, people, meetings) 2. link your notes together so they mirror how your brain actually thinks. 3. install obsidian cli so claude code can read your entire vault the relationships. 4. stop reexplaining projects every session. use reference files instead. 5. build custom slash commands: /context → load your full life work state /trace → see how an idea evolved over months /connect → bridge two domains you’ve been circling /ideas → generate startup ideas from your vault /graduate → promote daily thoughts into real assets 6. keep a strict rule: human writes the vault. agents read it, suggest, execute. 7. let claude aka clode surface patterns you’ve been unconsciously circling for years. 8. delegate from inside your notes. one sentence in obsidian → agent handles the rest. 9. treat writing as leverage.the more you write, the more context your agents have. 10. understand this:markdown files are the oxygen of llms. i really enjoyed seeing how to use obsidian thanks to @internetvin vin uses ai like a thinking partner wired into his life’s work. 99.99% of people won’t do this because it requires reflection setup. but once the vault exists, the agent stops being generic. it starts thinking in your voice. episode is live on @startupideaspod (more there) this one is different. send this tweet to a friend. im still processing how game changer obsidian claude code is, maybe you too watch

Here's 22 of the commands I am using with Obsidian and Claude Code with descriptions. I will turn this into something interactive soon so you can click the commands and then see the full prompts.

Vibe Coding Mobile Apps that COLLECT PAYMENTS (Complete Guide) In this video we: > Cloned an app that does $5M / Year w/ AI > Added API's (Gemini 3 and Nano Banana) > Added a @RevenueCat Paywall and tested it > Finished setup on App Store Connect Zero coding, built on the only app that can do this: @v_computer 00:00 Intro 00:43 The $5M App We're Cloning 02:14 First Prompt with Claude 4.5 Opus 02:49 Adding Gemini 3 API's and other Featurees 08:05 Testing Nano Banana Pro In App 09:04 Adding Payments to our App Starting the Flow 09:37 What is RevenueCat 10:11 Connecting RevenueCat 11:52 Prompting the Paywall with Products 13:52 Testing the Paywall 15:08 Altering the Paywall in RevenueCat 18:02 Shipping the App to the App Store 20:52 Additional Set up in App Store Connect 29:45 Testing the TEST FLIGHT Version 31:50 Final Thoughts

Oh my... this shouldn't be possible Creating a full stack mobile app in 493 seconds > Frontend > Backend with Database > AI integration (@NanoBanana Pro) > Payments integration (RevenueCat) > Sent to Apple App Store for review Zero Coding... whole process shown here. 00:00 Introduction 00:35 First Prompt for @v_computer 01:47 Adding Nano Banana 02:25 Adding Backend Database 02:58 Testing Nano Banana and Database 03:58 Fixing Error 04:56 Adding Payments 05:56 Generating Assets for Paywall 06:54 Testing the Paywall

Transformers v5's FINAL, stable release is out 🔥 Transformers' biggest release. The big Ws of this release: - Performance, especially for MoE (6x-11x speedups) - No more slow/fast tokenizers -> way simpler API, explicit backends, better performance - dynamic weight loading: way faster, and enabling: MoE now working w/ {quants, tp, peft, ...} We have a migration guide on the main branch; please take a look at it in case you run into issues. Come in our GH issues if you still do after reading it 😀

South Korea 🇰🇷 is now the clear #3 nation in AI — powered by the Korean National Sovereign AI Initiative there are now multiple Korean AI labs with near frontier intelligence. A key driver of this momentum is the Korean National Sovereign AI Initiative, a government-backed, nationwide competition that incentivizes domestic model development through a multi-stage elimination process. The initiative shortlists national champions, with winners receiving direct government funding and guaranteed access to large-scale GPU capacity. ➤ In August 2025, five organizations were selected: Naver, SK Telecom, LG Group, Upstage, and NC AI ➤ In the most recent round announced last week, the field narrowed to three: LG, SK Telecom, and Upstage. ➤ A fourth finalist is expected to be selected in the coming months as the evaluation process continues Generally, top Korean AI models tend to be open weights, and vary in size ranging from Motif‘s 12.7B Thinking model to LG’s 236B K-EXAONE. Other models, such as Korea Telecom (KT)’s Mi:dm K 2.5 Pro, are proprietary and developed with a focus on business integration with existing KT clients. Overview of major releases: ➤ LG | K-EXAONE - The current leader in the Korean AI race and a shortlisted model in the Korean National Sovereign AI Initiative. K-EXAONE is a 236B open weights model and scores 32 on the Artificial Analysis Intelligence Index. K-EXAONE performs strongly across various intelligence evaluations from scientific reasoning, instruction following, to agentic coding. However, this model has high verbosity, using 100 million tokens to run the Artificial Analysis evaluation suite ➤ Upstage | Solar Open - Another shortlisted model in the Korean National Sovereign AI Initiative. Solar Open is a 100B open-weights model and scores 21 on the Artificial Analysis Intelligence Index. Solar Open performs well in instruction following and has lower hallucination rate compared to peer Korean models ➤ Naver | HyperCLOVA X SEED Think - A 32B open weights reasoning model that scores 24 on the Artificial Analysis Intelligence Index. HyperCLOVA X SEED Think demonstrates strong performance on agentic tool-use workflows and scores highly in the Global MMLU Lite multilingual index for Korean, highlighting its potential usefulness in a primarily Korean language environment ➤ Korea Telecom | Mi:dm K 2.5 Pro - A proprietary reasoning model that scores 23 on the Artificial Analysis Intelligence Index. Mi:dm K 2.5 Pro sees strong performance in agentic tool-use. Mi:dm K 2.5 Pro currently has no publicly available endpoint. Instead, Korea Telecom primarily intends to package this model into product offerings and use this model to serve KT’s clients ➤ Motif | Motif-2-12.7B - A small open weights model that scores 24 on the Artificial Analysis Intelligence Index. Motif-2-12.7B performs well in long-context reasoning and knowledge, but is highly token intensive - using 120 million tokens to run the Artificial Analysis evaluation suite See Artificial Analysis for further details of the Korean models: artificialanalysis.ai/ Join our Discord community to discuss more: discord.gg/92EMmeRH

TRL v0.27.0 is out!! 🥳 It includes GDPO, the latest variant of GRPO for multi-reward RL ✨ GDPO decouples reward normalization to avoid reward collapse and improve per-reward convergence — developed by @NVIDIAAI @nbasyl_tw @SimonXinDong @GXiming @shizhediao et al.