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

MoE Architecture: why 21 billion is worth more than it seems

On July 6, 2026, Tencent made public Hy3 , a next-gen language model. The news spread fast in the global AI community, as shown by The Decoder . The most significant figure is not the total size of the model, but rather its architecture.

Hy3 uses an architecture Mixture-of-Experts (MoE) . All in all, the model packs a whopping 295 billion parameters in total. Yet, during any single inference run, only 21 billion parameters are actually fired up. This means the real computing cost is closer to a mid-sized model rather than a 295B giant.

The MoE principle is not new in the AI landscape. In fact, Google applied it successfully in Gemini models, and Meta explored it in some variants of LLaMA. However, Tencent brings this architecture into the open-source domain with unprecedented scale and transparency. Therefore, the release of Hy3 deserves a thorough analysis, especially for those evaluating the adoption of AI in marketing processes.

The numbers that define Hy3: performance and hallucination rate

Tencent claims that Hy3 matches models two to five times larger in terms of active parameters. That's a bold claim. So, it's a good idea to check it against available benchmarks.

The most interesting data concerns the hallucination rate . Hy3 records a value of 5.4%, which is cut in half compared to comparable benchmark 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 a direct impact on brand reputation.

According to research by McKinsey on the AI 2025 landscape , the reduction of hallucinations is one of the main obstacles to enterprise adoption of language models. Consequently, a 5.4% rate positions Hy3 competitively compared to much more expensive proprietary solutions.

On top of that, open-source code lets companies run the model on-premise. This means sensitive data never travels across third-party infrastructure. For industries like retail, finance, or B2B with strict compliance rules, this is a major structural perk.

Marketing use cases: where Hy3 finds concrete application

For Italian marketing managers, the practical question is simple: where can I use this model? The answer depends on your company setup and goals. Still, a few areas really stand out.

Content generation and SEO copywriting. Hy3 can support the production of SEO-optimized texts at scale. In particular, the reduction in hallucinations makes it more reliable for informational content and product pages. Anyone who wants to dive deeper into the implications for SEO content production will find this model to be an interesting tool to evaluate.

Campaign personalization. Efficient MoE models lend themselves to processing high-volume creative variations. Therefore, managers google ads campaigns or activities on Linkedin can explore workflows for the automated generation and testing of creative assets.

Analysis and reporting. Hy3 can be integrated into marketing data analysis pipelines. Similarly to other open-source models, it supports summarization, classification, and insight extraction tasks from structured and unstructured datasets. This integrates naturally with the services of Digital marketing which 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 to be a solid choice. Plus, the lower inference cost makes it much easier for SMEs with smaller budgets to get started.

Trade-offs to consider before adoption

No model is perfect. So, it is smart to weigh the main trade-offs of Hy3 before adding it to a tech evaluation.

• Hardware infrastructure: even with just 21B active parameters, running it locally takes proper GPUs. For many Italian small businesses, this means cloud spending or a specialized tech partner.
• Independent benchmark verification: performance data is reported 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: like with many models from China, performance in Italian might lag behind English or Chinese. This is something to test out hands-on for Italian marketing campaigns.
• Maintenance and updates: Open-source models need in-house skills or an outside partner to keep things running. On the flip side, proprietary APIs give you automatic updates but less control.

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

SHM Studio's perspective: open-source as a strategic lever, not just a tactical one

We at SHM Studio we're seeing a clear trend: the most advanced companies don't make a binary choice between proprietary and open-source AI. Instead, they build hybrid stacks, where models like Hy3 handle high-volume, low-sensitivity tasks, while enterprise solutions look after critical processes.

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

However, technology alone doesn't create value. As a result, 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 tech evaluation with hands-on design.

If you want to dive deeper into how open-source AI can fit into your marketing workflows, you can check out the SHM Studio services or check out our Blog for continuous updates on the AI landscape. For a direct comparison, the team is available through the contact page .

What nobody is saying: the geopolitical signal behind Hy3

There's a reading that goes beyond the tech stuff. Tencent's open-source release of Hy3 is not a neutral move. In fact, it fits into a broader strategy to position the Chinese AI ecosystem on the global market.

Making a competitive model available without licensing fees lowers adoption barriers for developers and companies worldwide. Therefore, it accelerates the spread of standards and architectures of Chinese origin. This competitive dynamism — which sees Tencent, Alibaba, and DeepSeek releasing open-source models at an increasing pace — is reshaping the balance of the global AI market.

For Italian decision makers, this means more choice and more complexity. In particular, it means having to evaluate not just technical performance, but also the implications of technological dependence, data governance, and alignment with European AI regulations. L’ digital architecture of companies will need to take these factors into account with growing attention over the next 12-24 months.