Tencent Releases Hy3: Open 295B MoE Model Under Apache 2.0 License

My take: Tencent releasing a model of this scale under Apache 2.0 says something important about where we are: high-performance AI is no longer exclusive to large proprietary APIs.

Hy3 is a Mixture-of-Experts (MoE) model: 295 billion parameters total, with only 21 billion activated per task. That means large-model quality at a fraction of the compute cost. In an independent evaluation with 270 domain experts on real workflows, Hy3 outperformed larger models on autonomous search (91.0 on DeepSearchQA) and tool orchestration (79.1 on MCP-Atlas), two areas that matter most for anyone building agents.

Does your company already have a strategy for evaluating open-source models as an alternative to proprietary APIs, or are you still fully dependent on a single vendor?