The leaked interface marks the first concrete look at Tencent's evolving AI strategy since the company previewed Hy4 during its latest earnings call. The model selection screen reveals a deliberate dual-track approach: maintaining proprietary control over flagship capabilities while integrating a leading open-source option — DeepSeek — to cover broader use cases with greater efficiency.
Hy4's positioning is notably more ambitious than its predecessor Hy3, with industry observers speculating that its "expert-grade" capabilities could stem from two likely technical directions. The first is a parameter-scale leap into the trillion-parameter range, bringing it on par with industry flagship models. The second is a deep reinforcement-learning (RL) post-training phase built on the Hy3 architecture, pushing the model's core intelligence ceiling higher without blindly expanding parameters. Either route points to a broader industry reality: competition among enterprise-grade LLMs has entered a stage defined by specialization and scenario-specific optimization.
Hy3, launched earlier this year, provides a useful baseline for understanding Hy4's trajectory. With 295B total parameters (21B active parameters and 3.8B in MTP-layer parameters), Hy3 has demonstrated strong performance in high-stakes tasks such as software development, office productivity, financial modeling, front-end design, and game production. In blind evaluations conducted internally by Tencent with domain experts performing real work tasks, Hy3 reportedly outperformed comparable models. This is telling: the yardstick for enterprise LLM quality is shifting from public benchmarks to expert-validated performance in real business scenarios.
From an industry perspective, major tech players are rapidly binding their proprietary LLMs to office AI agents. Hy4 is expected to fully take over as the foundational engine for WorkBuddy, unlocking synergies between model capabilities and productivity scenarios. This deep model-application coupling signals that enterprise LLMs are no longer standalone products, but rather "capability cores" embedded within workflows, collaboration tools, and agent ecosystems — redefining how enterprise AI is delivered.
For enterprises, choosing an LLM increasingly means choosing more than the model itself; it involves the surrounding toolchains, compute scheduling, and application ecosystem. No matter how impressive a model's raw performance, without a stable, schedulable computing-power foundation and mature engineering capabilities, it will struggle to deliver under real enterprise workloads. This aligns closely with the approach taken by StarWar Cloud