16-year-old trains 780B-parameter MoE model Ganymede

Sixteen-year-old Afonso, working under his small lab Merkium, trained Ganymede 1.1: a mixture-of-experts model with 780B total parameters and about 20B active per token, pretrained on roughly 5T tokens across 450+ H200 GPUs in three countries. It offers a 1M-token context and is free to use via chat and an OpenAI-compatible API at 30 requests per minute.

• MoE design: 780B total parameters, about 20B active per token
• Pretrained on roughly 5T tokens of web text and open-source code
• Trained on 450+ H200 GPUs across France, Portugal and the Netherlands
• 1M-token context, free access, rate limit of 30 requests per minute