Open always wins.

Is GPT-5 a good milestone to finally move on from calling everything "models" and start differentiating between chatbots, APIs, model weights & compare/benchmark accordingly? It makes no sense to compare chatbots outputs (which can be the result of dozens of different model weights) against a single model weight outputs, it's like comparing a car with an engine (book from @random_walker @sayashk)

BOOOOM! You can now run @OpenAI gpt-oss 20B natively in @GoogleColab T4 for FREE! 🔥 Powered by Transformers ⚡ The setup takes a bit since everything is bleeding edge, but once done it should work as expected Link to our cookbook in comments 👇

If you’re serious about mastering LLMs, agents, and AI systems start here. HuggingFace just released 9 elite-level AI courses for free ↓

Can’t catch a break even on my night off!

Light night reading

Both gpt-oss models are trending #1 and #2 among 2M models on @huggingface! 🤗 Thanks to the open-source AI community for your support since launch. We’re following discussions and will pop in when we can—feel free to ask questions, share ideas, and show what you’re building!

Great perspective and tips from @ClementDelangue on the early days of gpt-oss. Building strong foundations take time, and we’re excited to be working with inference providers and @huggingface to help these models perform at their best. Please keep your feedback coming!

Lots of conflicting takes about gpt-oss (yay open-source in the spotlight)! We’re powering the official @openai demo with HF inference providers thanks to @FireworksAI_HQ, @CerebrasSystems, @GroqInc and @togethercompute so we have a front-row seat of what’s happening. Something to remember: inference for new frontier open models isn’t easy, especially with a new format like harmony and the volume of interest that gpt-oss is getting out of the gate. Early spikes can temporarily affect quality, accuracy, and overall "vibes," particularly just 24 hours post-release when providers are racing against the clock with barely any sleep! Some advice to avoid forming the wrong opinions: - If you care about getting vibes as fast as possible, use a hosted setup that gives you a diversity of providers like HF inference providers ( and follow official standard guides ( - If you can, try it on your own hardware from the HF model pages ( which are usually good sources of truth with transformers and vLLM (keep in mind that you can make mistakes in generation there too) - Chill and take X hot takes with a grain of salt. Even if the first days can be confusing, the power of open-source is that there’s no cheating so we’ll uncover all the strengths and limitations of these models progressively! Personally, I think this is a landmark release by the @OpenAI team, the kind we will remember in 10 years because of its implications. Can’t wait to see what the future holds 🚀🚀🚀