The awkward secret behind the lightweight model boom over the past year is that most of the strongest Chinese releases are off-limits to a large portion of the businesses most interested in them. License terms excluding the European Union, the UK and South Korea meant that legal teams killed deployments before engineering teams had finished their assessments – not just for companies headquartered there, but for any business serving traffic to those regions. For IT teams who value open models, the tradeoffs are extraordinarily clear.
Tencent just removed that barrier. The company’s Hunyuan team released the full version Hy3A 295-billion-parameter Expert Mixture (MN) model with 21 billion active parameters and – unlike the original April release – shipped under permissive conditions. Apache 2.0 license. The reaction from the open model community was immediate, with researchers at X marking the license change as a real headline and claiming in a widely shared post that Tencent has become one of the leaders of open source if scores continue. Tencent says it will Two weeks free on OpenRouter.
It’s worth looking at the scores carefully – and they don’t all point in the same direction. But the more interesting story is where Tencent has chosen to lead: its reliability metrics and deployment economics are aimed squarely at production usage.
From first vision to product in ten weeks, shaped by 50 internal teams
Hy3’s April preview was the first model of Tencent’s revamped pre-training and reinforcement training infrastructure, shipping less than three months after its February overhaul. 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 model cardThe team gathered feedback from more than 50 product teams after a preview in late April, fixed task execution and interaction issues, and expanded the post-training pipeline.
The architecture is unchanged: 295B total parameters, 21B active per forward link via top-8 routing between 192 experts, a 3.8B-parameter multi-token prediction (MTP) layer for speculative decoding, and 256K context windows. It is behavior that changes. Tencent’s position is that the full release significantly outperforms similar-sized models and rivals flagship open-source models with two to five times the specs.
This "two to five times" The framework makes sense for where this model is headed, and it invites direct comparison with the current open weight coding leader, GLM-5.2.
Tencent’s blind test favors Hy3 over GLM-5.1, but GLM-5.2 still has the coding.
Tencent’s headline ranking is more of a blind human survey than a leaderboard. Arguing that public benchmarks don’t tell the full story, the company conducted a real-world workflow test with 270 professionals who collected 312 valid comparisons. work
The choice of opponent is important. Zhipu AI released GLM-5.2 in mid-June, and Tencent’s own benchmark plugin shows GLM-5.2 ahead of Hy3 in the entire agent encoding suite: SWE-bench Verified (84.2 vs. 78.0), SWE-bench Multilingual (83.0 vs. 75.8), Terminal-Bench 2.17.1, and WE171 vs. Wide (46.2 vs. 28.0). The blind test targeted the older model; the new one holds the coding crown.
Considering the sizes side by side, GLM-5.2’s coding figure is less surprising: GLM-5.2 is TN with approximately 744 billion parameters, versus Hy3’s figure of 295 billion total and 21 billion active. Tencent provides a model with less than half the parameters of the model it follows and about half the cost per token.
Hy3’s real wins lie elsewhere. In agent search, it ranks 84.2 on BrowseComp and 91.0 on DeepSearchQA — ahead of every open model on Tencent’s chart and competing with Claude Opus 4.8 and GPT-5.5. It leads the open field in tool orchestration (79.1 on the public MCP-Atlas suite), agent-harness evaluations such as ClawEval, and long-context search (73.4 on AA-LCR). Read together, the addition to GLM-5.2 offers a model with the best open-weight selection for search and tool-heavy agent workloads by forgoing repository-scale coding.
One caveat applies to both wins and losses: Almost all of the competitor numbers in Tencent’s supplement are listed as coming from Tencent’s own test races. Independent verification from indexes such as Artificial Analysis is still expected from the day of publication.
Confidence level: hallucination rates halved
Where the release is most interesting to enterprise buyers is the set of numbers Tencent chose to highlight instead of benchmarks. The model card reads less like a leaderboard announcement and more like a production reliability report.
In internal evaluations of real-world scenarios, Tencent says Hy3’s hallucination rate has dropped from 12.5% to 5.4% compared to the preview version, and overall error rates have dropped from 25.4% to 12.7%. do not confuse sources, do not fabricate information. Multi-loop behavior gets the same treatment: the drop rate in internal multi-loop tests dropped from 17.4% to 7.9%, and Tencent reported that the model’s score on the open MRCR long dialog benchmark rose from 42.9% to 75.1%.
Tencent also emphasizes consistency between agent scaffolds—noting the SWE-bench difference at several points, regardless of whether the model runs in Claude Code-style harnesses, Cline, or KiloCode. This is an underappreciated feature: enterprises rarely have control over which agent framework their teams standardize on, and a single-trailer model is a hidden integration cost. These are self-reported internal measurements and deserve the same skepticism as any vendor benchmark. But Tencent’s choice to put them at the forefront in all signals that it believes in its customer: teams fired by models that perform well and confidently prepare for production.
Placement math: Model 295B in a 744B world – on export compatible silicon
The reliability story is directly linked to economics, where the coding gap of Hy3 versus GLM-5.2 begins to look like an intentional trade-off rather than a loss.
GLM-5.2 is a TN with approximately 744 billion parameters, with approximately 40 billion active parameters per token; In FP8, its weights consume about 744GB alone, making 8x H200 nodes a practical minimum for serving production. Hy3 carries an FP8 footprint of under 300GB, at 295B total settings – less than half the memory, with about half of the active settings per token lowering the per-request calculation. For an organization that decides what to host on its own, it’s the difference between a lot of dedicated nodes with KV cache and storage space, and something more accessible.
There’s also a geopolitical wrinkle worth noting in the deployment guide: Tencent’s recommended service configuration targets Nvidia. H20-3e — A memory-enhanced variant of the H20, a GPU Nvidia designed specifically to comply with US export restrictions to China. Unlike GLM-5.2, there is no mention of Huawei or Ascend chips. In other words, the model is sized so that eight-chip Chinese companies can legally buy it and comfortably service it with full accuracy. Constraint-based design has a beneficial side effect for everyone: a model that works well with deliberately closed silicon works more smoothly on H100s, H200s and B200s available in Western data centers via standard. vLLM and SGLang Placements with MTP speculative decoding.
Add the Apache 2.0 license—no regional exclusions, no usage area restrictions—and the enterprise equation becomes clear. GLM-5.2 remains the obvious weight choice when encoding performance is the only criterion and the 8x H200 budget is available. Hy3 shows up everywhere: for search and tool-heavy agent workloads, reliability-sensitive applications, and organizations that want border-adjacent capability without border-scale infrastructure. The open question is whether Western enterprises will take Tencent’s model as a serious contender at all once the licensing hurdle is cleared – or will the next AI update settle the benchmark debate before it gets a chance to buy?
