Tencent Releases Hy3 295B parameter open weight model (huggingface.co)

Tencent has unveiled its Hy3 model, a substantial 295 billion-parameter mixture-of-experts (MoE) AI system featuring 21 billion active parameters. This release follows the Hy3 preview in April and incorporates extensive feedback from over 50 products, enhancing model performance and utility across a range of complex productivity tasks such as coding, financial modeling, and game development. Hy3 significantly outperforms similar models while being cost-effective, making it a notable contender in the current AI landscape. One of Hy3's standout features is its robust performance improvements across various operational metrics, with a marked reduction in issues related to hallucinations, commonsense reasoning, and multi-turn interactions. Notably, the model achieved a hallucination rate drop from 12.5% to 5.4% and reduced commonsense errors from 25.4% to 12.7%. These advancements make Hy3 particularly valuable for applications requiring high reliability and accuracy. With the ability to process long-context conversations effectively and enhanced fine-tuning capabilities, Hy3 is poised to set new benchmarks for AI model performance in multiple sectors while being accessible under the Apache License 2.0.