Tencent Hy3: The open-source MoE model challenging the giants

MoE Architecture: Why 21 Billion Is More Than It Seems

On July 6, 2026, Tencent made public Hi3, a next-generation language model. The news spread rapidly through the international AI community, as documented by The Decoder. The most relevant data is not the total size of the model, but its architecture.

Hy3 uses an architecture Mixture-of-Experts (MoE). In summary, the model has a total of 295 billion parameters. However, in each individual inference operation, only 21 billion parameters are activated. This means the actual computational cost is that of an average model, not a 295B giant.

The MoE principle isn't new to the AI landscape. In fact, Google has successfully applied it in its Gemini models, and Meta has explored it in some LLaMA variants. However, Tencent is bringing this architecture to the open-source domain with unprecedented scale and transparency. Therefore, the release of Hy3 deserves in-depth analysis, especially for those considering the adoption of AI in marketing processes.

Hy3's Defining Numbers: Performance and Hallucination Rate

Tencent claims Hy3 matches models two to five times its size in active parameters. This is an ambitious statement. Therefore, it is appropriate to contextualize it with available benchmarks.

The most interesting data concerns the hallucination rate. Hy3 has a value of 5.4%, which is half that of comparable reference models. For marketing applications—where factual accuracy is critical—this metric carries significant weight. In fact, a model that frequently hallucinates generates unreliable content, with direct repercussions on brand reputation.

According to the research of McKinsey on the AI Landscape 2025, reducing hallucinations is one of the main obstacles to the enterprise adoption of language models. As a result, a rate of 5.4% positions Hy3 competitively against much more expensive proprietary solutions.

In addition to this, opening the source code allows companies to run the model on-premise. Therefore, sensitive data does not travel over third-party infrastructure. For sectors like retail, finance, or B2B with compliance constraints, this is a structural advantage.

Use cases for marketing: where Hy3 finds practical application

For Italian marketing managers, the practical question is simple: where can I apply this model? The answer depends on the business context and objectives. However, some areas emerge clearly.

Content creation and SEO copywriting. Hy3 can support the production of search engine optimized texts at scale. In particular, the reduction of hallucinations makes it more reliable for informational content and product descriptions. Those who want to delve deeper into the implications for SEO content production will find this model an interesting tool to evaluate.

Campaign personalization. Efficient MoE models lend themselves to processing creative variants in volume. So, who manages Google Ads campaigns o activities on LinkedIn can explore automated creative asset generation and testing workflows.

Analysis and reporting. Hy3 can be integrated into marketing data analysis pipelines. Similar to other open-source models, it supports summarization, classification, and insight extraction tasks from structured and unstructured datasets. This integrates naturally with services digital marketing that include advanced analytical components.

Chatbots and internal assistants. Companies looking to deploy an AI assistant for their marketing team—without relying on external APIs—will find Hy3 a credible candidate. Furthermore, the reduced inference cost lowers the barrier to entry for SMEs that lack enterprise-level infrastructure budgets.

Trade-offs to consider before adoption

No model is without limitations. Therefore, it is useful to map Hy3's main trade-offs before incorporating it into a technology assessment.

• Hardware infrastructure: Even with only 21B active parameters, local deployment requires adequate GPUs. For many Italian SMEs, this implies a cloud investment or a specialized technical partner.
• Independent benchmark verification: The performance data are declared by Tencent. At the time of release, validation by the independent academic community was still ongoing. Therefore, it is advisable to wait for external reviews before making binding strategic decisions.
• Italian language As with many Chinese-language models, performance in Italian may be lower than in English or Chinese. This is an aspect that needs to be tested empirically in Italian marketing contexts.
• Maintenance and updates: Open-source models require in-house expertise or an external partner for ongoing management. In contrast, proprietary APIs offer automatic updates but less control.

According to Gartner, the operational maturity of open-source models has significantly increased, but the support gap compared to enterprise solutions remains a factor to consider. Similarly, the choice between open-source and proprietary is never just technical; it's also organizational.

SHM Studio's Perspective: Open-Source as a Strategic Lever, Not Just a Tactical One

We of SHM Studio We observe an established trend: the most advanced companies do not make binary choices between proprietary and open-source AI. Instead, they build hybrid stacks, where models like Hy3 handle high-volume, low-sensitivity tasks, while enterprise solutions oversee critical processes.

Hy3 fits into this scenario with an interesting profile. In fact, it combines computational efficiency, open source, and competitive performance. For marketing managers who are evaluating how to integrate AI into their processes—from SEO all AI consulting for marketing — this model deserves attention.

However, technology alone does not generate value. Consequently, the real differentiator is the ability to integrate these tools into concrete, measurable, and scalable workflows. This is precisely the work we do with our clients, combining technological assessment with operational design.

Those who want to delve deeper into how open-source AI can be integrated into their marketing processes can explore the SHM Studio services to consult our blog for continuous updates on the AI landscape. For direct comparison, the team is available through the Contact Us.

What no one tells you: the geopolitical signal behind Hy3

There is a reading that goes beyond the technical. Tencent's open-source release of Hy3 is not a neutral gesture. In fact, it fits into a larger strategy of positioning the Chinese AI ecosystem on the global market.

Making a competitive model available without licensing costs lowers adoption barriers for developers and companies worldwide. This accelerates the spread of Chinese-origin standards and architectures. This competitive dynamism — with Tencent, Alibaba, and DeepSeek increasingly releasing open-source models — is reshaping the global AI market landscape.

For Italian decision-makers, this means more choice and more complexity. In particular, it means having to evaluate not only technical performance but also the implications of technological dependency, data governance, and alignment with European AI regulations.’Digital architecture Companies will need to consider these factors with increasing attention over the next 12-24 months.