Munich🥨NLP Meetup Recap: Multilingual LLMs and AI Hardware at LRZ We spent yesterday at the Leibniz Supercomputing Centre (LRZ) diving into how specialized hardware is breaking the speed and language barriers of modern AI. Here are the highlights: 🎤 Michael Hoffmann & Jophin John (LRZ) — Llama-GENBA-10B: English, German, and... Bavarian? *The "Beergarten" Origin: What started as a casual chat in a beer garden turned into a serious research project: the first trilingual model for English, German, and Bavarian. *The Data Challenge: Low-resource languages are notoriously difficult. The team harvested Bavarian sentences from German newspapers and built a custom tokenizer to handle dialect nuances. *The Training Recipe: By using continuous pre-training followed by fine-tuning, they achieved state-of-the-art results for Bavarian, overcoming the total lack of standard benchmarks. *Efficiency at Scale: Trained on the Cerebras CS-2, the model saw 66 days of training with 35 MWh of energy consumption—a fascinating look at efficient foundation models. 🎤 Dmitry Gaynullin (Cerebras Systems) — Building Impossible Things Hardware Architecture: Unlike GPU clusters, the Cerebras system creates a static graph directly into its 850,000 cores. With "weight streaming," it handles massive models with the simplicity of a single node. Speed = Reasoning: Why aren't reasoning models everywhere? The main answer is speed. Dmitry showed how the Wafer-Scale Engine makes the "Thought-Act-Observe" loop significantly faster, making deep research agents practical for real-time use. 💻 Hands-on Workshop: Deep Research at Warp Speed We wrapped up with a hands-on session building our own deep research agents. The inference speed on the Cerebras hardware was truly eye-opening; seeing a reasoning agent move that fast changes how you think about AI workflows! A massive thank you to our hosts and speakers for making this happen! 🙏 From the Leibniz-Rechenzentrum team: Nicolay Hammer Michael Hoffmann Jophin John Leo Mak 🙏 From the Cerebras team: Alexander Mikoyan Dmitry Gaynullin And of course, thank you to the Munich🥨NLP community for the great questions and for braving the trip out to Garching! 🍕 #MunichNLP #LRZ #Cerebras #BavarianAI #LLM #GenerativeAI #WaferScale #AIInfrastructure #DeepResearch
-
-
-
-
-
+1