AMD Silo AI Expands Poro 2 Models with Enhanced Context and Math Capabilities
Lawrence Jengar Aug 06, 2026 08:52
AMD Silo AI releases Poro 2 8B Long with a 16x larger context window and Poro 2 8B Math Reasoning for multilingual AI development.
AMD Silo AI has unveiled two new open-source models in its Poro 2 large language model (LLM) lineup, aiming to push the boundaries of multilingual AI and digital sovereignty. The releases include Poro 2 8B Long, which features a dramatic 16x expansion of its context window to 131,072 tokens, and Poro 2 8B Math Reasoning, a math-focused variant designed for reasoning in Finnish and English. Both models and associated training resources are now available on Hugging Face, enhancing accessibility for developers and researchers.
The Poro family, developed in collaboration with TurkuNLP and trained on the LUMI supercomputer, has positioned itself as a leading example of adapting foundational LLMs for smaller, underrepresented languages. The original Poro 2 models, based on Meta’s Llama 3.1 architecture, focused on optimizing Finnish language performance while retaining strong multilingual capabilities. These new additions continue to build on the methodology first introduced with Poro 34B in 2024, which proved the value of multilingual pretraining for both language and coding tasks.
Breaking Barriers with Extended Context Windows
The standout feature of Poro 2 8B Long is its ability to process up to 131,072 tokens in a single interaction, a 16-fold increase over the original 8,192-token limit. This extended context window significantly enhances its utility in Retrieval Augmented Generation (RAG) workflows, making it capable of incorporating broader datasets or longer documents into its responses. For industries reliant on extensive text analysis, such as legal, academic, or enterprise search applications, this upgrade could be a game-changer for efficiency and accuracy.
Multilingual Math Reasoning: A First for Finnish
The Poro 2 8B Math Reasoning model is another key development, tailored for solving mathematical problems in both Finnish and English. While still in its preview phase and not yet production-ready, it addresses a significant gap in multilingual reasoning models. Current market solutions largely limit their reasoning traces to English or Chinese, leaving non-English speakers at a disadvantage when reviewing or verifying AI outputs. Finnish-speaking users now have the opportunity to analyze reasoning traces in their native language, a step toward more inclusive AI solutions.
Open-Source Access and Community Impact
Both models are released under open-source licenses, along with playbooks and training datasets that detail their development on AMD Instinct GPUs. By providing these resources, AMD Silo AI continues its focus on democratizing access to advanced AI tools, aligning with the goals of the OpenEuroLLM consortium. This pan-European initiative aims to strengthen the region’s digital sovereignty by fostering innovation in multilingual, open-weight AI models. The Poro 2 8B Long release, in particular, is a meaningful contribution to the ongoing effort to create AI systems tailored for European languages.
What to Watch for Next
AMD Silo AI has hinted at further advancements with the upcoming Poro 3 release later this year, which is expected to include expanded reasoning capabilities and new features for agentic workflows. For now, researchers and developers can explore the Poro 2 8B Long and Poro 2 8B Math Reasoning models on Hugging Face or consult the training playbooks available on the AMD ROCm blog. These latest releases highlight AMD's commitment to advancing multilingual AI and setting the stage for more autonomous and dynamic AI systems in the future.
Image source: Shutterstock.jpg)