Over the past year, the uncomfortable secret of the openweight boom has been that many of the strongest Chinese releases were out of reach for a large portion of the companies most interested in them. Licensing terms that excluded the European Union, the United Kingdom, and South Korea meant that legal teams canceled deployments before engineering teams finished their evaluations, not just for companies based there, but for any company serving traffic to those regions. For IT teams evaluating open models, the trade-offs are unusually explicit.
Tencent just removed that hurdle. The company’s Hunyuan team released the full version of hy3a 295 billion parameter Mix of Experts (MoE) model with 21 billion active parameters and, unlike the April preview, submitted it under the permissive apache 2.0 license. The reaction from the open model community was immediate: X researchers pointed to the license change as the real headline, and a widely shared post argued that if the scores hold, Tencent has just become one of the leaders of open source. Tencent says it will be like this free on OpenRouter for two weeks.
The results are worth examining, and they don’t all point in the same direction. But the more interesting story is what Tencent chose to lead: reliability and deployment economics metrics aimed squarely at production use.
From preview to product in ten weeks, made up of 50 internal teams
The April Hy3 preview was the first model of Tencent’s rebuilt pre-training and reinforcement learning infrastructure, shipping less than three months after the February rebuild. Chief AI scientist Shunyu Yao described the early open release as a deliberate move to gather feedback from developers and users before the official release, and Tencent says that’s exactly what happened. According to the model cardThe team collected feedback from more than 50 product teams after the late April preview, fixed issues in task execution and interaction, and expanded its post-training process.
The architecture is unchanged: 295 billion total parameters, 21 billion active per pass-through top-8 routing between 192 experts, a 3.8 billion parameter multi-token prediction (MTP) layer for speculative decoding, and a 256 billion context window. What changed is the behavior. Tencent’s positioning is that the full version significantly outperforms similarly sized models and rivals open source flagship models with two to five times the parameters.
That "two to five times" The framing makes sense for the purpose of this model and invites a direct comparison with the current leader in open weight coding, GLM-5.2.
Tencent’s blind test favors Hy3 over GLM-5.1, but GLM-5.2 still has encryption
Tencent’s headline evaluation is a blinded human study rather than a league table. Arguing that public benchmarks don’t tell the full story, the company conducted a blind test with 270 experts from all disciplines working on real-world workflows, collecting 312 valid comparisons, in which Tencent reports that Hy3 scored 2.67 out of 4 versus 2.51 for GLM-5.1, with the clearest advantages in frontend development, CI/CD, and data and storage work.
The choice of opponent is important. Zhipu AI launched GLM-5.2 in mid-June, and Tencent’s own benchmark appendix shows GLM-5.2 ahead of Hy3 in virtually the entire agent coding suite: SWE-bench Verified (84.2 vs. 78.0), SWE-bench Multilingual (83.0 vs. 75.8), Terminal-Bench 2.1 (81 vs. 71.7), and DeepSWE by a wide margin (46.2 vs. 28.0). The blind test focused on the previous model; the newer one retains the encoder crown.
GLM-5.2’s coding lead is less surprising once you consider the sizes side by side: GLM-5.2 is roughly an MoE of 744 billion parameters with around 40 billion active parameters per token, versus Hy3’s 295 billion total and 21 billion active. Tencent is introducing a model with less than half the parameters (and almost half the calculation per token) of the next one.
Hy3’s real victories lie elsewhere. In agent search, it scores 84.2 in BrowseComp and 91.0 in DeepSearchQA, ahead of all open models in Tencent’s table and competitive with Claude Opus 4.8 and GPT-5.5. It leads the field in tool orchestration (79.1 in the MCP-Atlas public suite), in agent utilization evaluations such as ClawEval, and in long context retrieval (73.4 in AA-LCR). Read together, the appendix suggests a model that is arguably the best open weight option for tool- and search-heavy agent workloads, while granting repository-scale coding to GLM-5.2.
One caveat applies to both wins and losses: Almost all of the competitor numbers in Tencent’s appendix are marked as coming from Tencent’s own testing. Independent verification, of indices such as Artificial Analysis, is still pending at the time of publication.
The Reliability Argument: Hallucination Rates Halve
Where the launch becomes more interesting for enterprise buyers is the set of numbers that Tencent chose to emphasize rather than benchmarks. The model card looks less like a rating advertisement and more like a production reliability report.
In internal evaluations in real-world scenarios, Tencent says Hy3’s hallucination rate fell compared to the preview version from 12.5% to 5.4%, and common sense error rates fell from 25.4% to 12.7%; improvements he attributes to detailed data cleaning and training constraints built around an explicit behavioral pattern: respond when substantiated, indicate when evidence is missing, do not combine sources, do not fabricate data. Multi-turn behavior gets the same treatment: The emission rate in internal multi-turn tests fell from 17.4% to 7.9%, and Tencent reported that the model’s score on the MRCR open long dialogue benchmark jumped from 42.9% to 75.1%.
Tencent also emphasizes consistency across agent scaffolds: it reports SWE bank variance at a few points, whether the model is run within Claude Code, Cline, or KiloCode-style harnesses. That’s an underrated property: companies rarely control which agent framework their teams standardize on, and a model that only works in one harness has a hidden integration cost. These are self-reported internal measurements and deserve the same skepticism as any vendor benchmark. But the decision to put them front and center in everything indicates who Tencent believes its customer is: teams that have been burned by models that demo well and build confidently in production.
The mathematics of implementation: a 295 billion model in a world of 744 billion, with export-grade silicon
The history of reliability connects directly to economics, and this is where the Hy3 vs. GLM-5.2 coding gap starts to look like a deliberate trade-off rather than a loss.
GLM-5.2 is an MoE of about 744 billion parameters with about 40 billion active parameters per token; on FP8, its weight alone consumes approximately 744 GB, making an 8x H200 node the practical minimum for production service. Hy3, with 295 billion total parameters, occupies an FP8 space of less than 300 GB: less than half the memory, with about half the parameters active per token, resulting in less computation per request. For an organization deciding what to self-host, that’s the difference between a high-spec node and something much more attainable, with space left over for KV cache and batch processing.
There is a geopolitical issue in the deployment guide that is also worth noting: Tencent’s recommended service configuration targets Nvidia. H20-3e – the memory-enhanced variant of the H20, the Nvidia GPU designed specifically to comply with US export restrictions to China. Unlike the GLM-5.2, there is no mention of Huawei or Ascend chips here. In other words, the model is sized so that eight of the chips that Chinese companies can legally buy can comfortably serve with complete precision. That constraint-driven design has a convenient side effect for everyone else: a model that works well with deliberately limited silicon works even more comfortably on the H100, H200, and B200 available in Western data centers, via standard. vLLM and SGLang Implementations with MTP speculative decoding.
Add the Apache 2.0 license (no regional exclusions or field of use restrictions) and the business equation becomes clear. GLM-5.2 remains the open weight option when encoding performance is the only criterion and an 8x H200 budget is available. Hy3 makes its case everywhere else: tool- and search-heavy agent workloads, reliability-sensitive applications, and organizations that want edge-adjacent capabilities without edge-scale infrastructure. The open question is whether Western companies, now that the licensing barrier is gone, will treat a Tencent model as a serious candidate, or whether the upcoming Artificial Analysis update settles the benchmark debate before acquisitions have a chance.
