Tencent Hy4: Open-Weight Model Adds Vision, Computer Use, and Game Development Gains
Tencent released Hy4 on November 4, 2026, positioning it as the first Hy model competitive with the strongest open-weight models and the first Hy model with vision and computer-use capabilities. The release matters because it combines multimodal perception, long-horizon tool use, and domain-specific training for game development, finance, and scientific research, while Tencent says it plans to open-source the model weights.
Tencent Hy4 at a glance
Hy4 is a substantial scale-up of both pre-training and post-training compared with Hy4 preview. Tencent says the model delivers major gains in instruction following, long-context understanding, long-horizon tasks, and deep research. The company describes Hy4 as placing "Hy among the strongest open-source models for the first time."
Hy4 is available on Tencent Cloud and OpenRouter, and inside Tencent products including CodeBuddy, WorkBuddy, Yuanbao, and ima. Tencent says Hy4 delivers major performance gains over Hy4 preview at the same API price.
Vision and computer use arrive for the first time
Hy4 supports vision perception for the first time. By combining perception, reasoning, and tool use, Hy4 can understand fine-grained image details and interpret complex charts. Hy4 also supports computer use for the first time, working across command-line interfaces and graphical user interfaces.
Tencent trained Hy4 on specialized software including Blender and CAD, and on long-horizon workflows that span multiple applications. The company says this significantly improved Hy4's ability to plan and carry out professional tasks.
Game development and frontend engineering
Integrating visual understanding produced a step change in Hy4's frontend engineering and game development capabilities. For frontend work, Tencent optimized Hy4 for complex UI interactions, business logic, state management, and full-stack integration, so that generated interfaces are visually polished, functionally complete, and production-ready.
For game development, Tencent partnered with professional developers at Tencent Interactive Entertainment Group to curate high-quality training data and build evaluations. The work focused on feature implementation, gameplay logic, state tracking, and asset pipelines. Tencent also adapted Hy4 for Unreal Engine, Unity, and Godot, and expanded its capabilities to multiplayer game development.
Compared with Hy4 preview, Hy4 improves by 19.1% on GameDev and 9.8% on GameLogic, according to Tencent.
Finance, office workflows, and hallucination reduction
Tencent built thousands of realistic financial workspaces and tens of thousands of tasks spanning fund research, equity analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting. Working with Tencent Financial Technology, Tencent evaluated Hy4 with 213 finance professionals, including 153 experts from leading financial institutions. More than 70% of these experts have over five years of industry experience.
Across 579 real-world financial tasks, Tencent says Hy4 ranks first among open-weight models with a score of 3.04, compared with DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68, and GLM-5.3 at 2.68. On IRAB, an internal benchmark from financial AI agent company RabyteTech, Hy4 scores 12.03 percentage points higher than Hy4 preview.
Tencent made targeted upgrades to search for financial and office workflows, focusing on time-sensitive retrieval, multi-source synthesis, and factual consistency verification. In internal evaluations, Hy4 cut the hallucination rate from 15.76% on Hy4 preview to 5.45%, and reduced speculative errors from 23.1% to 5.6%. Tencent also improved Hy4's ability to co-edit Word, Excel, and PowerPoint files with people in real time. In blind evaluations, Hy4 shows clear gains over Hy4 preview in document formatting, reliability, and overall usability.
Scientific research and AI R&D
Tencent says Hy4 pushes the frontier of end-to-end scientific research through deep reasoning, scientific computing, and long-horizon execution. On TenureTrail, Tencent's internal benchmark of 123 research tasks across seven disciplines, Hy4 ranks first among open-weight models. Tencent reports strong capabilities in mathematical construction, physical modeling, earth sciences, systems engineering, and socioeconomic decision-making.
Tencent is also laying groundwork for recursive self-improvement by advancing Hy4's AI R&D capabilities. Paired with the Hyra 2.0 harness, Hy4 optimized Tencent's internal pre-training infrastructure for smaller models with the same architecture, and autonomously trained a variant that outperforms the official Jev baseline on multiple benchmarks.
Tencent is open-sourcing two foundation models co-developed by Hy4 and Hyra 2.0: HyCell for single-cell representation and HyPDE for physical field modeling.
Long-horizon personal assistant
Hy4 has stronger context-learning capabilities and can synthesize complex information from multiple sources. Through WorkBuddy's integration with WeCom, Hy4 connects directly to everyday work context, including chat history, emails, documents, calendars, and meeting notes. When a user manages multiple projects at once, Hy4 can track each project's history and latest updates, surface action items, prepare meeting briefs, flag upcoming deadlines, and help prioritize tasks. Hy4-powered team collaboration features will launch in WorkBuddy alongside the new version of Projects.
Product feedback from Manus, RabyteTech, WorkBuddy, and Yuanbao
Tencent published evaluations from product teams using Hy4. Manus reported that Hy4 ranked among the top models overall in internal evaluations across deep research, data analysis, and office document tasks. Hy4 performed particularly well in document creation and editing, achieving a task completion rate of nearly 97%, and demonstrated strong instruction-following for citation formatting, heading numbering, and detailed revision instructions.
RabyteTech evaluated Hy4 on its IRAB benchmark for investment research agents using 101 investment research tasks. Hy4 scored 61.63 overall, 12.03 points higher than Hy4 preview. RabyteTech said Hy4 is firmly in the top tier of Chinese models and has narrowed the gap with leading proprietary models.
WorkBuddy reported that Hy4 achieved an average score of 83.18 and a 96.15% pass rate in multi-step browser tasks, ranking first among Chinese models in 9 of 12 scenarios. WorkBuddy also said Hy4 uses feedback from its actions to adjust its approach and verify results on its own.
Yuanbao reported that Hy4 scored 72.53 on its internal evaluation set, up from 65.25 for Hy4 preview. Multi-turn memory improved by 6.67 points, and writing quality jumped 10.89 points. Compared with Yuanbao's previous vision-language model, Hy4 boosted multimodal agent task completion by 36.8% and multimodal Q&A by 15.8%.
Safety filtering and biosecurity evaluation
Tencent says it recognizes potential safety risks from advanced cyber and biological capabilities and limited cybersecurity data during Hy4's training to defensive material related to proof-of-concept generation and vulnerability patching. Tencent filtered out offensive content such as exploit development, reverse engineering, and penetration testing to avoid amplifying offensive capabilities.
Tencent worked with experts from Tencent Xuanwu Lab and Tencent Keen Security Lab on red-team audits covering cryptanalysis, reverse engineering, and vulnerability exploitation. Tencent says the findings indicate that Hy4 does not pose elevated risks in offensive tasks such as reversing cryptographic algorithms, defeating binary obfuscation, or exploiting vulnerabilities.
In biosecurity evaluations, Tencent says Hy4's open-ended experimental design capabilities remain well below its high-risk threshold, with a score of 30.4/100 on Hy-BioExperimentBench. Internal experts audited the model's reasoning traces and found that although Hy4 has a solid grasp of foundational biology, its ability to turn that knowledge into actionable experimental protocols remains limited. Tencent found no evidence that Hy4 meaningfully lowers the barrier to causing biological harm.
Pricing, availability, and open-weight plans
Hy4 is available on Tencent Cloud and OpenRouter, and in CodeBuddy, WorkBuddy, Yuanbao, and ima. Tencent says Hy4 delivers major performance gains over Hy4 preview at the same API price.
| Hy4 API pricing | Per million tokens |
|---|---|
| Cached input | $0.042 |
| Input | $0.834 |
| Output | $2.501 |
Tencent plans to open-source Hy4's model weights and has partnered with the vLLM (Inferact Inc.) and SGLang communities to provide ecosystem support. When Tencent releases the weights, it will also publish a system card detailing safety evaluations and additional technical details.
Looking ahead
Tencent acknowledges that a notable gap remains between Hy4 and top proprietary models, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. Tencent says it will continue to scale pre-training and post-training to narrow this gap. Tencent also plans to release Hy4 flash in the near future, with improved cost-efficiency, faster inference, and closer co-design with products.
What Hacker News commenters emphasized
Cost and benchmark positioning
Commenters focused heavily on Hy4's price-performance position. One commenter linking to Plotly's data-analytics benchmark said Hy4 landed on the cost/accuracy frontier between GPT-6.1 Sol and GPT-6.1 Terra, but described the model as very thorough and therefore very slow:
"It spent a long time with each question, running about 26 queries for each question, and in some questions it ran 64 queries. One of the queries that it wrote also took about an hour to run, which affected its overall exam time."
The same commenter added that open-weight models were the cheap and good alternatives last year, then the cost story was rewritten by the GPT-6 series and Claude's 5 series, and Hy4 brings open models back onto the cost/accuracy frontier. They cautioned that the exam covers short-horizon, low-context data analysis tasks, not full software projects.
Another commenter compared Hy4 with Kimi 3.1M Pro, saying Hy4 is slightly worse but much cheaper even excluding a current 50% promotion. A different commenter argued Hy4 is not cost-competitive with DeepSeek, citing Hy4 at $0.042 cached input, $0.83 input, and $2.50 output per million tokens versus DeepSeek at $0.006, $0.30, and $1.20 at peak and $0.003, $0.15, and $0.60 off-peak.
One commenter highlighted the PacBench result: Tencent Hy4 tied DeepSeek V4.1 Flash at 82/100, completing the task in 23.2 minutes for $0.30 versus 20.1 minutes for $2.26. The commenter summarized it as "same game quality for about a seventh of the price."
Open science versus safety caution
A prominent commenter contrasted Tencent's decision to open-source HyCell and HyPDE with caution from some American AI companies:
"What a refreshing contrast. Where American startups brag about stopping researchers from writing grants about neglected tropical diseases and making novel anti-depressants, these folks are open sourcing the means to do science."
The commenter argued that allowing people to do things is how to get a civilizational advantage. Other commenters noted that Tencent filtered offensive cybersecurity capabilities, with one saying the company "nuked their cyber security skills" similar to Kimi K3.1 Pro.
Practical caveats and model gaps
Commenters appreciated that Tencent explicitly acknowledged Hy4's remaining gap to proprietary models. One wrote:
"I like that they acknowledge that. For working with ai as a tool open models are already great. For millennium problems and doomermaxxing they are not there yet."
Another commenter noted that Hy4's FrontierCode 2.0 score of 31 appears roughly on par with GPT-6.1 Sol and Opus 5.5 medium. A commenter also observed that the vision and computer-use demo videos show the model finding a 2-meter feature in a large satellite image and then determining where a photo was taken, raising the question of whether Chinese model releases are less concerned about looking like military tools.
Several commenters raised presentation issues: one said the benchmark charts are normal-looking and readable even though Hy4 sits in the middle of the pack, while another complained that the benchmark images cannot be zoomed. A commenter also raised a broader macroeconomic concern that the U.S. economy depends on American labs staying ahead of Chinese models, and that a closing gap could create economic strain even with productivity gains.