IBM Research

We traveled to the Basque Country this week for a wonderful event unveiling Europe’s first IBM Quantum System Two, which is now operational at the IBM-Euskadi Quantum Computational Center in San Sebastian, Gipuzkoa, Spain: https://lnkd.in/edsde576. This milestone positions the Basque Country as a global leader in quantum computing, and is just the latest example of how our partners at the Basque Quantum Initiative (BasQ) are continually demonstrating their commitment to advancing science, technology, and innovation. Real quantum computers can speed up scientific discovery of quantum algorithms and applications, and with an IBM Quantum System Two powered by two 156-qubit Heron chips, the European quantum community has gained a powerful new tool for scientific discovery. Thank you to everyone who joined us for this week’s event and helped make it such a success. It was a pleasure to talk with Imanol Pradales Gil about the future and the opportunities for us to partner and build together.

Today we officially unveiled the very first IBM Quantum System Two in Europe - in beautiful San Sebastien, Spain. Already operational and proving to be one of our most powerful computers to date, this installation marks another technical milestone in the journey to bring useful quantum computing to the world, and help grow a thriving, talented, global quantum ecosystem to push the boundaries of science and innovation. Congratulations to the IBM Quantum & Ikerbasque + BasQ teams on this momentous occasion. Thank you for so kindly welcoming us to your beautiful Basque Country. 🇪🇸 Check out the annct: https://lnkd.in/ejMKjz2k More on the IBM-BasQ partnership here: https://lnkd.in/e_k7q-yB IBM Research Jay Gambetta Horacio Morell Ana Paula De Jesus Assis Alessandro Curioni Rajeev Malik Mikel Diez Parra Jeannette Garcia Paola Ortega Hernández Jason Sinclair Cristina Sanz Fernández Cristina Rodriguez Diaz Dr. Voica Radescu Chris Nay

The next generation of quantum hardware in Europe is here. We’re proud to announce the installation of Europe’s first IBM Quantum System Two, now operational at the IBM-Euskadi Quantum Computational Center in San Sebastián, Spain: https://ibm.co/6041BKZQ3 - La próxima generación de hardware cuántico en Europa ya está aquí. Nos enorgullece anunciar la instalación del primer IBM Quantum System Two de Europa, ya operativo en el Centro de Computación Cuántica IBM-Euskadi en San Sebastián, España.

𝗡𝗼𝘁 𝗮𝗹𝗹 𝘁𝗼𝗸𝗲𝗻𝘀 - 𝗼𝗿 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 - 𝗮𝗿𝗲 𝗰𝗿𝗲𝗮𝘁𝗲𝗱 𝗲𝗾𝘂𝗮𝗹. Inference scaling methods like Best-of-N, Majority Voting, and Beam Search all boost model accuracy, but each trades off latency and token cost differently. Some are parallel and fast; others are sequential and deliberate. Over the summer, I had the joy of mentoring two exceptionally talented MIT students Jenny Y. Huang, Mehul Damani, alongside my incredible colleagues Ramón Fernández Astudillo, Yousef El-Kurdi, to investigate how to select the right method and allocate the right amount of compute per query - trading off 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆, 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, and 𝘁𝗼𝗸𝗲𝗻 𝘂𝘀𝗮𝗴𝗲. The result is our initial study, 𝘓𝘢𝘵𝘦𝘯𝘤𝘺 𝘢𝘯𝘥 𝘛𝘰𝘬𝘦𝘯-𝘈𝘸𝘢𝘳𝘦 𝘛𝘦𝘴𝘵-𝘛𝘪𝘮𝘦 𝘊𝘰𝘮𝘱𝘶𝘵𝘦 (https://lnkd.in/eu5fEtRp) - exploring how models can “𝘵𝘩𝘪𝘯𝘬 𝘧𝘢𝘴𝘵” when they can, and “𝘵𝘩𝘪𝘯𝘬 𝘩𝘢𝘳𝘥𝘦𝘳” only when they must. Let's face it - the future isn’t about thinking less or thinking more - 𝗶𝘁’𝘀 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗦𝗺𝗮𝗿𝘁 😉 Excited for where this line of work will take us next! Go #sustainableAI! Go IBM Research! 🫶

𝐔𝐧𝐯𝐞𝐢𝐥𝐢𝐧𝐠 𝐨𝐮𝐫 𝐧𝐞𝐰 .𝐭𝐢𝐧𝐲 𝐚𝐧𝐝 .𝐬𝐦𝐚𝐥𝐥 𝐏𝐫𝐢𝐭𝐡𝐯𝐢 𝐚𝐧𝐝 𝐓𝐞𝐫𝐫𝐚𝐌𝐢𝐧𝐝 𝐄𝐚𝐫𝐭𝐡 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐭𝐢𝐨𝐧 𝐦𝐨𝐝𝐞𝐥𝐬 Size matters, but bigger is not always better when it comes to generative AI. As many examples have shown, radically reducing your model’s size doesn’t just make it cheaper and faster to run, easier to fine-tune, or less resource-hungry. More importantly, it enables deployment in far wider and often constrained, settings, where a good share of real impact lies. That’s why we’ve extended our flagship open source model families, #Prithvi and #TerraMind, with 𝐭𝐢𝐧𝐲 and 𝐬𝐦𝐚𝐥𝐥 variants, with great performance and all the good features like Thinking in Modalities, etc Thanks to this work, we’ve achieved model 𝐬𝐢𝐳𝐞 𝐫𝐞𝐝𝐮𝐜𝐭𝐢𝐨𝐧𝐬 𝐮𝐩 𝐭𝐨 120 𝐭𝐢𝐦𝐞𝐬 with a 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐝𝐫𝐨𝐩 𝐨𝐟 𝐥𝐞𝐬𝐬 𝐭𝐡𝐚𝐧 10%. And that supporting a frozen encoder design. Imagine placing the encoder on a satellite in orbit, and then uploading tiny decoder heads (e.g. just 1–2 MB for classification tasks) as needed to add new capabilities. That’s a big deal given uplink constraints. In collaboration with Unibap Space Solutions AB, we’ve already tested this approach on space-grade hardware processing live Sentinel-2 imagery. But this goes far beyond satellites. These new models enable advanced AI to run efficiently on modest hardware, bringing pervasive intelligence to environments with limited bandwidth or compute power. Developers will appreciate the simplicity, no massive GPU clusters required. Fine-tuning can now be done on a laptop. And of course, we’re keeping it 𝐨𝐩𝐞𝐧 (truly open source, not only the weights), and available to the entire community with truly permissive licences on Hugging Face. More on our blog: https://lnkd.in/e4vf7_Us Hugging Face Prithvi: https://lnkd.in/ejczfrX8 Hugging Face TerraMind: https://lnkd.in/egW3V3rC NASA - National Aeronautics and Space Administration European Space Agency - ESA Jülich Supercomputing Centre (JSC) German Aerospace Center (DLR) Rahul Ramachandran Manil Maskey Sujit Roy, Ph.D. Nicolas Longépé Giuseppe Borghi 🌍🛰️🇪🇺 Clem Delangue 🤗 Paolo Fraccaro Johannes Jakubik Benedikt Blumenstiel Campbell Watson Romeo Kienzler Zoë Gathercole Alessandro Curioni IBM IBM Research #AI4Science ⬇️ Below a video showing TerraMind.tiny using inference to find elephants in images and Prithvi-EO-2.0 Tiny to classify land use from images... links/code in the blog