See you next Tuesday at @NeurIPSConf. We’ll be at 𝗸𝗶𝗼𝘀𝗸 𝟳! Make sure to stop by, discover our latest research, and pick up a sticker!

To celebrate 𝗙𝗹𝘂𝘅 𝟮 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗮𝘁 𝗳𝘂𝗹𝗹 𝘀𝗽𝗲𝗲𝗱 after our collaboration with @bfl_ml and @replicate, we’re launching the Flux 2 x Pruna Contest on X! 🎨 Create an image using any Flux 2 model on BFL or Replicate 🏆 Prizes: 𝟭𝟱𝟬€ 𝗽𝗲𝗿 𝗰𝗮𝘁𝗲𝗴𝗼𝗿𝘆 +

@plagiat_os @bdsqlsz @replicate @PrunaAI with the assist https://t.co/fI9KpugGRS

Z-model from @AlibabaGroup is now running is less than 2 seconds on @replicate optimized by @PrunaAI ! We've accelerated inference so you can run Z-model faster, cheaper, and greener. And this is just the beginning, even more speed-ups are coming soon! 👉 Try it on Replicate:

Pruna is built … • with Pretzels & Croissants 🥨 🥐 • in Paris & Munich & in the metaverse 🇩🇪 🇫🇷 👾 So, we moved our offices online to @gather_town and things have been great! Cool things we’ve seen happen: • more engaging and fund global meetings • great excuses for

Flux 2 from @bfl_ml is out at maximum inference speed on @replicate with @PrunaAI optimization! - The novelty: Stunning 4MP quality with new state-of-the-art realism and high-precision control with up to 8 reference images. This is a huge milestone for Black Forest Labs! - The

⏰ 𝗦𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝗮𝘁 𝟲:𝟬𝟬 𝗽𝗺 𝗖𝗘𝗧! Learn from @chrismdesa from Cornell University how simple linear error feedback unifies two key challenges in modern ML: efficient training and efficient inference. In today’s webinar, Chris will present: • 𝗚𝗿𝗮𝗕

🚀 Excited to see you at our 𝗔𝗜 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 𝗠𝗲𝗲𝘁𝘂𝗽 𝗶𝗻 𝗠𝘂𝗻𝗶𝗰𝗵 on 𝗗𝗲𝗰𝗲𝗺𝗯𝗲𝗿 𝟭𝟬𝘁𝗵 𝗮𝘁 𝟲:𝟯𝟬 𝗣𝗠 𝗖𝗘𝗦𝗧! 📅 Agenda • 6:30 PM: Doors open • 7:00–8:00 PM: Talks featuring, • 8:00–9:00 PM: Hang out with likeminded developers & Food 🍕 👉

🏋️‍♀️ A proper warmup is important, especially for compiled models, but Pruna can now do without it. torch.compile can accelerate your model even with LoRA. But the warmup and hot swapping can be rather slow. 🐌  Our improvements: 🔧 Portable Compilation: Save compiled artefacts

We kicked off the @dotaiconf with a burst of creativity, designing new 𝗣𝗿𝘂𝗻𝗮 𝗔𝗜 𝘀𝘁𝗶𝗰𝗸𝗲𝗿𝘀. Developers, researchers, and ML engineers teamed up to create merch design that captures what we care about: 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁, 𝘀𝗰𝗮𝗹𝗮𝗯𝗹𝗲, 𝗮𝗻𝗱 𝗳𝘂𝗻 𝗔𝗜. Some

🎓 Learn from @chrismdesa how simple linear error feedback unifies two key challenges in modern ML: efficient training and efficient inference. In this webinar, on Nov. 24th, Chris will present: • 𝗚𝗿𝗮𝗕 (𝗚𝗿𝗮𝗱𝗶𝗲𝗻𝘁 𝗕𝗮𝗹𝗮𝗻𝗰𝗶𝗻𝗴) — a new way to select training

⚡️ AI is slow, expensive… and bad for the environment! At @dotaiconf, I shared how energy utilization interact with AI models — and how we can make it much more efficient. 💡What drives AI progress? Energy utilization drives AI breakthroughs like transformer architectures,

2 seconds for 4MP is wild. Great work from the @PrunaAI team!

While everyone is waiting for Flux 2 from @bfl_ml, we built an ultra-fast 4 MP model based on Flux-Schnell. It only takes 2 seconds per generation. You can try it now on @replicate.

🎓 Learn from @chrismdesa how simple linear error feedback unifies two key challenges in modern ML: efficient training and efficient inference. In this webinar, on Nov. 24th, Chris will present: • 𝗚𝗿𝗮𝗕 (𝗚𝗿𝗮𝗱𝗶𝗲𝗻𝘁 𝗕𝗮𝗹𝗮𝗻𝗰𝗶𝗻𝗴) — a new way to select training

Do you want to know what we can do for models on your costly H100s?! From March to May, we did some pretty cool things on @replicate • FLUX runs in less than a second (0.5s) • Hidream models series run 1.3x to 2.5x faster • Wan 2.2 runs on 8 GPUs instead of 1, making it

Huge thank you to everyone who joined us yesterday to create a new merch design for @PrunaAI ! The design will be launched soon, stay tuned 👀 If you’re at the @dotaiconf today, come say hi. And don’t miss @Bertrand_Charp talk at 2:50 PM, see you there! 🚀

We want to congratulate @bria_ai_ on releasing FIBO, their frontier image generation model, which was trained on structured JSON for precise, controllable generation. Some cool things - x5 faster and powered by our optimisation expertise - JSON prompts? Yes, it gives you more

🎃 Nothing scarier than slow and costly models. Joking but not joking. Happy Halloween! Pruna is ready for it. Boooo. Spooky!

Get your laptops ready and put on your optimisation gloves for a hands-on session in Paris on Nov 5 (18:30–21:00) with @dotaiconf! We're running a live session on model optimisation, including quantisation, compilation, LoRA tricks, and more. Get ready to get hands-on! 👩‍💻 Ideal