Tencent Drops Hy3, a 295B Open-Weight Coding Agent Slashing Hallucinations in Half

Tencent just dropped the full release of Hy3, its latest open-weight model from the Hunyuan team, and it is now available for free on OpenRouter and OpenCode. This is the production-ready successor to the Hy3 Preview that shipped in late April, rebuilt from the ground up with a focus on agent reliability, hallucination reduction, and real-world coding performance.

A sparse giant that punches above its weight

Hy3 is a 295B-parameter Mixture-of-Experts model that activates only 21B parameters per token. That distinction matters a lot in practice. The total parameter count is a marketing headline; 21B active is what shapes latency and per-token cost. Sparse activation lets a ~300B model punch like a much larger dense stack on agent benchmarks, without activating every weight.

The architecture is built around efficiency at scale:

• 192 experts with top-8 routing, so only 8 experts fire per token, keeping compute low.
• A Multi-Token Prediction (MTP) layer that predicts several tokens at once for faster decoding, enabled through speculative decoding in both vLLM and SGLang.
• A 256K token context window, supporting full-repo ingestion and long document tasks.
• Weights published under Apache License 2.0, with an FP8 quantized variant also available.

What changed from Preview

Tencent says this release is better than the April preview after gathering feedback from 50 internal product teams. The team focused on three specific failure modes that matter most when running models in production pipelines:

• Tool-call stability: The team fixed baseline stability issues that broke agents. Invalid calls that trigger infinite loops dropped. Hy3 also generalizes across agent scaffoldings, with SWE-Bench Verified accuracy variance across CodeBuddy, Cline, and KiloCode staying within 4%.