February 2025 Summaries

Koyeb has announced the availability of Tenstorrent Instances on its serverless platform, providing rapid access to Tenstorrent's Wormhole multi-chip solution aimed at enhancing AI model inference performance. This collaboration introduces two new instance types, TT-N300S and TT-Loudbox, which leverage Tenstorrent's Tensix Processor and open-source TT-Metalium SDK to offer developers scalable and high-performance AI infrastructure without the need for extensive infrastructure management. The integration allows for instant development environments, fast NVMe volumes, continuous deployment, and built-in observability, all accessible through one-click applications that simplify deployment and connectivity. By enabling on-demand access to these advanced AI resources, the partnership underscores Tenstorrent's commitment to fostering innovation and providing developers with the tools to experiment with and adopt their technology.

Globula, a geolocation-based augmented reality game, integrates real-world exploration with multiplayer role-playing and storytelling, and has chosen Koyeb to manage its infrastructure as it scales. Initially reliant on a single powerful server, Globula faced challenges such as single points of failure, connectivity risks, and downtime during updates. The decision to switch to Koyeb was driven by the need for a reliable, scalable solution that includes features like built-in autoscaling, multi-region deployments, continuous deployment, and advanced networking. Using Koyeb's platform has allowed Globula to efficiently manage sudden traffic surges, reduce costs through Scale-to-Zero, and ensure low latency and high availability worldwide. The infrastructure improvements have enabled Globula to focus on enhancing the game and aim for a target of over 200,000 daily players, while also benefiting from faster deployments and uninterrupted player experiences.

Open-source large language models (LLMs) are increasingly competing with proprietary models in various natural language processing tasks, offering advanced AI capabilities without the high costs and restrictions associated with closed models. However, deploying these open-source models at scale remains a challenge, which can be addressed through serverless GPU solutions that simplify infrastructure management. The blog post reviews several leading open-source LLMs as of early 2025, such as DeepSeek-R1, Mistral Small 3, and Qwen 2.5 Coder, highlighting their strengths in reasoning, conversational AI, and code generation, respectively. These models, available under open-source licenses, are noted for their competitive performance against proprietary models and their adaptability for specific tasks through fine-tuning. The post also emphasizes the ease of deploying and optimizing these models using platforms like Koyeb, which provide scalable and cost-effective solutions for running AI applications.