Tencent releases Hy3, a free 295B open model

Tencent released Hy3 on July 6. The model is a mixture-of-experts design: 295 billion total parameters, 21 billion active per token. That ratio cuts inference cost substantially compared to a dense model of equivalent nominal size. Context window is 256K. License is Apache 2.0 for both use and derivative works.

Tencent tested the model against 50 internal product teams before the public release. Hallucination rate dropped from 12.5% to 5.4% compared to the April preview, and commonsense errors fell from 25.4% to 12.7%. Average daily token consumption across Tencent products has increased 20x since the April preview. The model is integrated into WorkBuddy, Yuanbao, Marvis, and ima.

Where it runs

Hy3 is available on Hugging Face and ModelScope immediately. On OpenRouter, the route is tencent/hy3:free at $0 per token through July 21, after which a paid tier will apply. It is also supported in Cline, OpenCode, Kilo, and Cherry Studio on day one.

How it benchmarks

VentureBeat’s testing found Hy3 beats GLM-5.2 on most tasks at roughly half the active parameter count, while losing on coding benchmarks. The model targets long-document work, large codebases, and multi-step reasoning. For those categories, 256K context and 21B active parameters offer a cost-effective alternative to closed frontier models.

What this means

The free window ends July 21. Hy3 sitting at $0/token on OpenRouter is an opportunity to run eval sets against your current closed-model baseline at no cost. If your use case is document processing, summarization, or multi-step analysis rather than code generation, Hy3 is worth the 10 minutes it takes to add the route to your test harness.

The guide on whether open models are good enough for operators covers the evaluation framework. The Kimi K2.7 release brief and the GLM-5.2 leaderboard brief document the other open models now competing at this tier.