Tencent Hy3 beats larger rivals at half the size with Apache 2.0 license, reshaping open AI

Tencent Hy3 Removes Geographic Barriers With Apache 2.0 License

Tencent Hy3 arrived on July 6, 2026, marking a turning point for enterprises that had been locked out of China's strongest open models 1

1

. The Hunyuan team released the full version of this 295-billion-parameter model under the permissive Apache 2.0 license, reversing the restrictive terms from April's preview that excluded the European Union, United Kingdom, and South Korea 1

1

. Legal teams that previously killed deployments before engineering teams finished evaluations can now move forward. The open-source MoE model gives companies wide commercial freedom to self-host, fine-tune, and build global products without geographic restrictions 2

2

. Researchers on X singled out the license change as the real headline, with one widely shared post arguing that Tencent has become one of the leaders of open source 1

1

.

Mixture-of-Experts Architecture Cuts Inference Costs

The AI model operates as a Mixture-of-Experts system with 21 billion active parameters per forward pass via top-8 routing across 192 experts, despite its 295 billion total parameters 1

1

. Each prompt only passes through part of the network, which cuts inference costs and serving expenses 2

2

. The architecture includes a 3.8-billion-parameter multi-token prediction layer for speculative decoding and a 256K context window designed to handle long documents, large codebases, and multi-step work 1

1

2

2

. Tencent will offer the model free on OpenRouter for two weeks 1

1

.

Reduced Hallucination Rates After 50 Internal Team Reviews

Tencent Hy3 evolved from preview to product in ten weeks, shaped by feedback from more than 50 product teams after the late-April preview 1

1

. Chief AI Scientist Shunyu Yao framed the early open release as a deliberate move to gather feedback before the official version 1

1

. The team fixed issues in task execution and interaction, then scaled up its post-training pipeline 1

1

. By Tencent's numbers, hallucinations dropped from 12.5% to 5.4%, and commonsense errors fell from 25.4% to 12.7% 2

2

. These reliability metrics aim squarely at production use rather than benchmark chasing 1

1

.

Agentic Search and Tool Orchestration Lead, Coding Trails GLM-5.2

Tencent Hy3 excels in agentic search, posting 84.2 on BrowseComp and 91.0 on DeepSearchQA, ahead of every open model in Tencent's table and competitive with Claude Opus 4.8 and GPT-5.5 1

1

. It leads the open field on tool orchestration with 79.1 on the public MCP-Atlas set, on agent-harness evaluations like ClawEval, and on long-context retrieval with 73.4 on AA-LCR 1

1

. Tencent ran a blind test with 270 experts across disciplines working on real-world workflows, collecting 312 valid comparisons, in which Hy3 scored 2.67 out of 4 against GLM-5.1's 2.51, with advantages in frontend development, CI/CD, and data and storage work 1

1

.

However, Zhipu AI's GLM-5.2 keeps the coding crown. Tencent's own benchmark appendix shows GLM-5.2 ahead across the entire agentic coding suite: SWE-bench Verified at 84.2 versus 78.0, SWE-bench Multilingual at 83.0 versus 75.8, Terminal-Bench 2.1 at 81 versus 71.7, and DeepSWE by a wide margin at 46.2 versus 28.0 1

1

. GLM-5.2 is roughly a 744-billion-parameter MoE with around 40 billion active parameters per token, nearly double Hy3's active compute 1

1

. Nearly all competitor numbers in Tencent's appendix come from Tencent's own test runs, and independent verification from indices like Artificial Analysis is still pending 1

1

.

Production Deployments Signal Enterprise Readiness

Tencent Hy3 is already deployed in Tencent WorkBuddy, Yuanbao, WeChat assistants, and Path of Exile: Advent, making it more substantial than a research demo built mainly to post benchmark scores 2

2

. The model joins Zhipu AI, DeepSeek, Alibaba's Qwen, and Mistral AI in putting pressure on closed models from OpenAI and Anthropic 2

2

. For enterprises weighing an open-weight choice for search-and-tool-heavy agent workloads, Hy3 presents a compelling option with deployment economics that favor efficiency over raw parameter count 1

1

.