Tencent has launched Hy4 at a set price of about $0.83 per million input tokens and $2.50 per million output tokens, targeting autonomous software operation.
About two months after releasing a preview, Tencent rolled out the finished language model with multimodal image understanding and computer-use capabilities. The launch pushes enterprise AI competition beyond chat interfaces and directly into software manipulation, but the operational economics remain tightly controlled: model weights are currently closed, restricting deployment to Tencent Cloud, OpenRouter, and native applications like the Yuanbao chatbot.
Last July, Tencent released Hy3 with 295 billion parameters under the Apache 2.0 license, but the shift to Hy4 centers on practical workflow execution. What does it actually cost an enterprise to integrate Tencent's agent architecture against alternative models, and can its real-world performance justify a proprietary cloud lock-in?
Enterprise Tooling and Financial Benchmarks
Tencent specifically trained Hy4 to handle sequential application tasks and professional software, including CAD programs and Blender. For gaming development pipelines, Tencent adapted the model for engines such as Unreal, Unity, and Godot, while testing configuration tasks inside its own production studios.
Tencent states that Hy4's hallucination rate during web research dropped from nearly 16 to about 5 percent in internal tests.
Specialized corporate workflows served as the primary benchmark during evaluation. In an evaluation conducted with over 200 finance professionals across nearly 600 real-world tasks, Tencent reported that Hy4 ranked ahead of local models. Financial platform Comein placed Hy4 on par with Anthropic's Claude Opus 5.5 in second place among models it tracks. Beyond external workflows, Tencent utilized Hy4 alongside its Hyra 2.0 agent system to optimize internal training pipelines, independently training a version of US startup TypeSafe AI's Jev decision-making model.
How do these vendor-reported numbers hold up once the model leaves curated test environments?
Benchmark Disparities and Autonomy Limits
Independent evaluations reveal clear technical boundaries when external verification tools are applied. On the Artificial Analysis Intelligence Index—which measures knowledge work, coding, science, and general knowledge—Hy4 scored 35 points, landing above the median of 30 points.
When it doesn't know an answer, it makes one up 89 percent of the time instead of admitting the gap.
The critical operational risk surfaces when network tools are disconnected. On the AA-Omniscience factual accuracy test without web search assistance, Hy4 answered roughly half of the questions correctly from its base training data. When encountering unknown information without live retrieval, it fabricated answers 89 percent of the time rather than flagging an informational deficit.
Tencent positions Hy4 as an enterprise engine capable of operating complex office and engineering software at competitive inference rates. Yet outside web-connected retrieval, an 89 percent hallucination rate on unknown data and closed model weights leave corporate buyers dependent on cloud APIs, balancing promised desktop automation against unverified autonomous reliability.
What came before
Betting on specialized software pipelines is a logical pivot, yet the engine quickly falters whenever offline environments cut off real-time external retrieval. In factual accuracy tests without a live search engine, the model fabricated answers in 89 percent of cases when facing unknown data.
Our stories: Tencent Hy3: 21B Active Parameters Outperform Massive LLMs (July 7).