Tilde Trains 30B+ Balto-Slavic LLM on AMD Instinct ™ GPUs

Trained on the LUMI supercomputer using AMD Instinct™ MI250X GPUs, TildeOpen LLM — a large language model focused on Balto-Slavic languages — represents a major milestone in building open, locally relevant AI infrastructure that helps facilitate security and regulatory compliance, increases data ownership, and allows for fine-tuning towards tailored applications in multiple sectors.

AMD Silo AI has effectively demonstrated the training of multilingual LLMs on AMD hardware with its Poro and Viking model families tailored for Nordic languages. Following in similar footsteps, TildeOpen LLM, built by Tilde, one of Europe’s leading language technology companies, represents an important step towards advancing multilingual AI and digital sovereignty.

High-Performance, Open-Weight Foundation Model on AMD Instinct™ MI250X

TildeOpen LLM is an open-weight 30+ billion parameter foundation model, trained to support Balto-Slavic languages spoken by over 155 million people worldwide. Focusing on both scale and downstream applications, the model is large enough to support sophisticated reasoning and multilingual tasks while being small enough to be deployed locally, enabling data privacy.

“We were positively surprised by how simple it was to start training on AMD Instinct GPUs. Having previously used other hardware, we expected some level of compatibility issues with this new software-hardware configuration. Turns out our concerns were unnecessary as the transition was seamless.” says Toms Bergmanis, AI Researcher at Tilde.

This frictionless adoption highlights the maturity of the AMD ROCm™ open software stack for large-scale AI training. ROCm software provides the tools for programming AMD GPUs, from low-level kernels to high-level end-user applications, and has support for popular ML frameworks, such as PyTorch, JAX and TensorFlow.

AI Sovereignty Through Local Deployment

Unlike many general-purpose LLMs hosted in proprietary clouds, TildeOpen LLM can be fine-tuned for various applications and is designed to be deployed locally — enabling AI sovereignty for enterprises, governments, and research institutions through data control, privacy and security, model behavior, and infrastructure.

“TildeOpen LLM, along with AMD Silo AI Poro and Viking model families, are examples of critical AI infrastructure components, trained on AMD hardware. These models contribute to and enable an open ecosystem of AI software,” said Peter Sarlin, CEO & Co-founder at AMD Silo AI. “Developing local AI capabilities, models, and products is a critical advancement for sovereignty, an advancement we are committed to support by working with leading AI capabilities globally.”

Fine-Tuning and Public Release

TildeOpen LLM has completed its foundational training phase and is now published to Hugging Face. After fine-tuning, adapted models will also be published on Hugging Face under open licenses — further advancing access to powerful LLMs for Balto-Slavic applications.

AMD Silo AI is committed to enabling companies like Tilde to build critical digital infrastructure components. Whether it's powering real-time translation, private document summarization, or regional AI agents, the TildeOpen LLM is built to serve real-world needs — in Balto-Slavic languages, contributing to regional AI sovereignty.

_{© 2025 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, AMD Instinct, ROCm, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective owners. Certain AMD technologies may require third-party enablement or activation. Supported features may vary by operating system. Please confirm with the system manufacturer for specific features. No technology or product can be completely secure.}