On July 6, 2026, Tencent’s Hunyuan team released the full version of Hy3, a 295-billion-parameter Mixture-of-Experts model built for enterprise agentic workloads. The technical specifications are compelling, but the real headline for most enterprise AI teams is a single word in the license field: Apache.
The April 2026 preview of the same model shipped under a restrictive Tencent Hy Community License that excluded the European Union, the United Kingdom, and South Korea. For any company serving users in those markets, legal review killed deployment before engineering got to run evals. The July full release removes that barrier entirely. Apache 2.0 allows unlimited commercial use, modification, and redistribution anywhere in the world, with no royalties and no field-of-use restrictions.
Why License Terms Broke the Adoption Math
Open-weight models have reshaped the enterprise AI landscape over the past 18 months. But “open weights” and “open for enterprise” are not synonyms. Many of the strongest Chinese-origin models, including earlier Hunyuan and Qwen releases, shipped with licenses containing geographic exclusions. This effectively made them unavailable to multinationals, European-based teams, and any company with global compliance requirements.
The pattern created an unusual dynamic: models that could outperform closed-source alternatives on specific benchmarks, at a fraction of the compute cost, sat unused by the organizations most equipped to evaluate and deploy them. Apache 2.0 changes that equation. There are no geographic exclusions, no field-of-use restrictions, and no royalty clauses. The model can be deployed on any cloud, in any VPC, in any jurisdiction.
For enterprise AI teams that have been building on OpenAI or Anthropic APIs primarily due to licensing clarity, Hy3 is now a legitimate evaluation target, particularly for workloads where vendor lock-in or data residency constraints have been a persistent bottleneck. This development is directly relevant to the broader pattern of open-weight models competing for enterprise inference workloads at scale.
What Hy3 Is Under the Hood
The architecture is a Mixture-of-Experts design with 295 billion total parameters and 21 billion active parameters per forward pass, using top-8 routing across 192 experts. A 3.8-billion-parameter multi-token prediction (MTP) layer enables speculative decoding for lower latency. The context window is 256,000 tokens.
Tencent rebuilt its pre-training and reinforcement learning infrastructure from scratch in February 2026. The April preview was the first model from that rebuilt stack. Between April and July, the team collected feedback from more than 50 internal product teams (including Tencent Docs, CodeBuddy, WorkBuddy, Yuanbao, and the QQ ecosystem), fixing reliability issues and scaling up post-training data quality and diversity, per the official model card on Hugging Face.
The memory footprint matters for deployment decisions. At FP8 precision, Hy3 occupies under 300GB, meaning a 4x H100 node handles initial serving with room for KV cache and batching. GLM-5.2, the current open-weight coding leader, runs roughly 744 billion total parameters and requires approximately 744GB in FP8: an 8x H200 node at minimum. For enterprises evaluating self-hosting, that difference translates to roughly half the hardware cost before you run a single benchmark.
The Benchmark Picture: Where Hy3 Wins, Where It Concedes
Tencent chose to lead the release with reliability metrics rather than leaderboard positions. The company ran a blind evaluation with 270 domain experts, generating 312 valid comparisons on real work tasks. Hy3 scored 2.67 out of 4, beating GLM-5.1 (the predecessor model) at 2.51, with the clearest wins in frontend development, CI/CD tasks, and data and storage workflows. The choice of GLM-5.1 rather than GLM-5.2 as the comparison target is notable: GLM-5.2 shipped in mid-June 2026 and Tencent’s own appendix shows it ahead across the coding suite.
The performance story breaks down cleanly by workload type:
| Benchmark | Tencent Hy3 | GLM-5.2 | Workload Type |
|---|---|---|---|
| BrowseComp | 84.2 | n/a | Agentic search |
| DeepSearchQA | 91.0 | n/a | Agentic search |
| MCP-Atlas | 79.1 | n/a | Tool orchestration |
| SWE-bench Verified | 78.0 | 84.2 | Agentic coding |
| SWE-bench Multilingual | 75.8 | 83.0 | Agentic coding |
| Terminal-Bench 2.1 | 71.7 | 81.0 | Agentic coding |
| DeepSWE | 28.0 | 46.2 | Agentic coding |
Source: Tencent Hy3 GitHub repository. Competitor numbers from Tencent’s internal evaluation runs. Independent third-party verification is pending as of July 7, 2026.
The coding gap is real and Tencent did not obscure it. GLM-5.2 leads every repository-scale coding benchmark by a meaningful margin. Where Hy3 leads is in search-and-tool agent pipelines. Its BrowseComp score of 84.2 is competitive with Claude Opus 4.8 and GPT-5.5, per VentureBeat’s analysis of the release. On MCP-Atlas, the tool orchestration benchmark, it leads open-weight models in Tencent’s evaluation at 79.1.
For GTM and operations teams building agents that retrieve, filter, and synthesize information across tools rather than rewriting code repositories, this profile maps directly to their workload. A model that orchestrates MCP tool calls reliably and handles long search-then-synthesize pipelines is more valuable than one that tops SWE-bench if the actual deployment never touches a git repository.
One important caveat: VentureBeat noted that nearly all competitor numbers in Tencent’s benchmark appendix are marked as coming from Tencent’s own evaluation runs. Independent verification from indices like Artificial Analysis is still pending as of publication. Teams should treat these numbers as directionally useful, not definitive, until third-party results appear.
Reliability Metrics: The Production-Grade Story
For enterprise deployments, the more actionable numbers are not peak benchmark scores but failure rates across extended operation. Tencent reported the following improvements from the April preview to the July full release:
| Metric | Hy3 Preview (April 2026) | Hy3 Full (July 2026) |
|---|---|---|
| Hallucination rate (internal) | 12.5% | 5.4% |
| Commonsense error rate | 25.4% | 12.7% |
| Multi-turn issue rate | 17.4% | 7.9% |
| MRCR long-dialogue benchmark | 42.9% | 75.1% |
Source: Tencent Hy3 model card and official GitHub repository.
The multi-turn performance improvement is especially relevant for enterprise agent deployments. The MRCR long-dialogue score nearly doubling (42.9 to 75.1) indicates the model can track intent and context across extended interactions without the coreference and drift issues that affect many models in long agentic loops.
In Tencent’s WorkBuddy integration, the model handles complex agent workflows of up to 495 sequential steps with a reported success rate exceeding 99.99%. The Time to First Token dropped by 54% and end-to-end response time dropped by 47% compared to the April preview. These figures come from Tencent’s internal production deployment; external validation does not yet exist. But the metric profile aligns with what enterprise operations teams need: not a model that answers individual questions well, but one that completes long autonomous task chains without failing halfway through.
Available Now: API, Self-Hosting, and Framework Support
Hy3 is available through three routes today:
Tencent Cloud TokenHub API: 1 RMB per million input tokens, 4 RMB per million output tokens, with a 0.25 RMB cache-hit rate. At current exchange rates, this sits well below closed-model alternatives for most agentic workloads. The API is available at hunyuan.tencent.com and listed on OpenRouter.
OpenRouter: Free API access for two weeks from the July 6 launch. Standard pricing follows after the free window closes. For teams running quick evals, this is the fastest path to a comparison.
Self-hosted weights on Hugging Face: Available at tencent/Hy3 under Apache 2.0. Compatible with vLLM and SGLang inference frameworks. Agent frameworks including OpenClaw, OpenCode, KiloCode, Cline, and CherryStudio have already integrated support.
For enterprise teams with data residency requirements or outbound network restrictions, self-hosting on internal infrastructure is now legally straightforward for the first time with a Hunyuan-class model. Combined with the reduced hardware footprint compared to GLM-5.2, the barrier to a serious internal evaluation has meaningfully dropped.
What Enterprise AI Teams Should Do This Week
The right path depends on what your agents actually do.
If your primary workload is autonomous code generation at repository scale, GLM-5.2 remains the stronger open-weight option on current benchmarks, though it requires substantially more infrastructure and is available under MIT license from Zhipu AI.
If your workloads involve agentic search, multi-step information retrieval, MCP tool orchestration, multi-turn customer or sales interactions, or any workflow combining document processing with API calling, Hy3 warrants an evaluation run this week. The benchmark profile aligns with the kinds of pipelines GTM, RevOps, and operations teams are actually building, not with the academic coding benchmarks that tend to drive leaderboard coverage.
For teams already running enterprise agentic models like Claude Sonnet 5 in production, Hy3 is worth a comparison on your specific workloads, particularly if cost or data residency has been a constraint. The license, the memory footprint, and the agent-search benchmark profile give it a distinct position in the open-weight tier.
The broader signal here is that the gap between frontier closed models and enterprise-viable open-weight models continues to close. Licensing was the last major structural barrier for many global organizations. For Hy3, that barrier is now gone.
If you want to understand how a model like Hy3 fits into an enterprise agentic deployment strategy, including workload selection, infrastructure sizing, and where specialized models still apply, book a conversation with the Enera team.