Tencent Hy3: the open-source MoE model taking on the giants

Updated October 12, 2026

A 295-billion-parameter model that costs as much to run as a 21-billion one: Tencent released Hy3 on July 6, 2026, and cut the hallucination rate in half compared with comparable competitors. If you're an Italian SME eyeing AI with one eye on the budget, it's time to understand what really changes.

Is Hy3 really worth evaluating for my company, or is it just for big tech?

Hy3 is relevant for Italian SMEs too: running costs are those of a mid-sized model, not an enterprise system, and the open code lets you install it on your own servers. The right question isn't “can I afford it?” but “do I have the skills to run it?”

The issue isn't just technical. In our AI and process automation cluster, we've often described how the most advanced companies build hybrid stacks: open-source models for high-volume tasks, enterprise solutions for critical processes. Hy3 fits right into that pattern.

What is the MoE architecture, and why does it change the math?

MoE (Mixture-of-Experts) is a system in which the model activates only some of its components for each individual operation. The result: large-model performance at a mid-sized model's cost.

Imagine an office with 100 specialists. To answer a marketing question, you don't call all 100: you involve 7 or 8, the ones who know that topic. The cost is that of 7-8 people, not 100. Hy3 works the same way.

The model has 295 billion total parameters, but it activates only 21 billion in each operation. This mechanism is called sparse activation. A useful comparison: in the Mixtral-8x7B model, analyzed by PatSnap in a study of over 30 MoE patents, each token accesses 47 billion total parameters, but only 13 billion take part in the actual computation, a 3.6x reduction in active compute.

Google applies it in its Gemini models, and Meta has explored it in some LLaMA variants. Tencent brings this architecture to open-source with unprecedented scale and transparency, as documented by The Decoder.

Do Hy3's numbers hold up in an honest comparison?

Tencent claims Hy3 matches models two to five times larger in active parameters. The most verifiable figure is the hallucination rate: 5.4% versus 12.5% for comparable reference models, according to the company's launch statements.

The hallucination rate measures how often an AI model makes up false information and presents it as true. If you produce marketing content, product sheets or reports, a model that hallucinates often is a real problem: it generates unreliable text that needs constant manual review.

According to the data published by Tencent Hunyuan at the model's launch, Hy3 has a rate of 5.4%, half that of comparable models, which stood at 12.5%. It's a significant improvement, but the figures are self-reported by Tencent: independent validation by the academic community was still under way at the time of release. It's wise to wait for external reviews before making binding strategic decisions.

A second factor to consider is the Italian language. As with many Chinese-developed models, performance in Italian may be lower than in English or Chinese. This needs to be tested empirically before putting Hy3 into a content production workflow for the Italian market.

What's happening to enterprise AI costs: has anything really changed?

Yes: the average cost per AI token in the enterprise has dropped 67% in a year. Open source and routing across multiple models are the two main drivers behind the decline.

According to an analysis published by Open Source For You in May 2026, the average cost per million tokens fell from $18.40 in Q1 2025 to $6.07 in Q1 2026. That's a 67% drop in twelve months.

Multi-model routing means automatically using the cheapest model that fits each individual task: a lightweight model handles simple questions, and complex ones get escalated. With its cost-performance profile, Hy3 is a good fit for this kind of architecture.

This trend makes open-source models like Hy3 even more appealing for Italian SMEs that have so far held off because of budget. For a comparison with other quality open-source models, read our article on Mistral AI as an open-source alternative to OpenAI.

Is the Italian AI market really ready, or are we still behind?

The market is growing fast (€1.8 billion in 2025, up 50% from the previous year), but the gap between large companies and SMEs is huge: 71% of large companies have already launched AI projects, compared with just 8% of small and medium-sized ones.

According to the Artificial Intelligence Observatory at Politecnico di Milano, the Italian AI market reached €1.8 billion in 2025. The growth is real, but very unevenly distributed.

Data from the Digital Innovation in SMEs Observatory at the same university show that 76% of Italian SMEs have neither invested in AI nor plan to. Only 7% have launched structured training programs for their staff. The gap isn't just technological: it's cultural and organizational.

One positive sign: according to ISTAT data on Italian SMEs updated to 2026, AI adoption among small and medium-sized businesses more than tripled in two years, rising from 5% in 2023 to 16.4% in 2025. The direction is right, but the pace is still slow compared with large companies.

Where can an Italian company actually apply Hy3?

The most immediate areas are content production, marketing data analysis and internal chatbots. The starting point is always the specific use case, not the model.

Three illustrative scenarios to show where Hy3 can create real value. The numbers are calculation examples, not real client data.

If you run Google Ads campaigns or LinkedIn activities, Hy3 can support the generation and automated testing of high-volume creative variants. To learn how Chinese open-source models are changing the options available to Italian companies, read our analysis of Chinese open-weight AI models and their impact on businesses.

If you want to deploy an AI assistant for your marketing team without relying on external APIs, Hy3 is a credible candidate. Its lower inference cost lowers the barrier to entry compared with traditional enterprise solutions.

What open-source tools are out there, and how do they compare with Hy3?

Hy3 isn't the only quality open-source model: the market offers alternatives with different profiles in terms of language, cost and ease of deployment. The right choice depends on your use case and the infrastructure you have available.

Choosing between open source and proprietary is never purely technical: it's organizational too. An open-source model requires in-house expertise or an external partner for ongoing management. Proprietary APIs offer automatic updates but less control over your data. For a deeper look at how companies are ceasing to "rent" artificial intelligence, our article on open-source AI and proprietary on-premise models offers the full picture.

What risks and constraints should I know about before adopting Hy3?

There are three main risks: hardware infrastructure, still-partial independent verification of benchmarks, and ongoing model management. None is insurmountable, but all need planning.

Even with only 21 billion active parameters, local deployment requires adequate GPUs. For many Italian SMEs, that means a dedicated cloud investment or a specialized technical partner. It's not a prohibitive cost, but you need to factor it in before you start.

There's also a reading that goes beyond the technology. Tencent's open-source release of Hy3 is part of a broader strategy to position the Chinese AI ecosystem on the global market. Making a competitive model available with no license fees lowers the barriers to adoption for developers and companies worldwide, speeding up the spread of Chinese-origin standards and architectures. For Italian decision makers, this means more choice and more complexity: you need to weigh not only technical performance, but also the implications of technological dependence and alignment with European AI regulations.

How do you get started: what steps should you follow to evaluate and adopt Hy3?

The path has six steps, from defining the use case to monitoring in production. Each step has a clear owner and a measurable output.

What are the most common mistakes people make when approaching an open-source model like Hy3?

The most common mistake is starting from excitement about the model instead of the problem to solve. The second is underestimating ongoing running costs.

• Choosing the model before the use case. "Hy3 is powerful, let's find something to do with it" is the fastest way to waste budget and time.
• Ignoring hidden costs. The model is free, but GPUs, cloud, maintenance and team training are not.
• Trusting only the vendor's benchmarks. Tencent's data on Hy3 is promising, but independent validation is still partial. Test on your real use case.
• Skipping the Italian test. Performance in Italian can differ significantly from English. A test on real samples is a must before go-live.
• Forgetting governance. An AI model that produces content or analysis for your company falls under the European AI Act. Ignoring it is not an option.
• Not training the team. According to data from Politecnico di Milano, only 7% of Italian SMEs have launched structured AI training programs. Without training, even the best model delivers mediocre results.

How can you tell if Hy3 is really working for your company?

You need measurable indicators, not impressions. There are four KPIs to track: output quality, speed, cost per task and manual review rate.

So, is Hy3 worth it for my company?

It's worth it if you have a high-volume use case (content, analysis, classification), an in-house technical team or a reliable partner, and you want to keep data on-premise for compliance or GDPR reasons.

A middle ground: you're an SME with no in-house IT but a trusted technical partner. You can start with a limited pilot project, measure the results and decide whether to scale. The entry cost has dropped enough to make testing sustainable.

Not yet if you haven't identified a specific use case, you don't have suitable infrastructure and you can't count on technical skills in-house or outside. In that case, a managed proprietary API is still the safest choice.

How we see it at SHM

For us, AI should be integrated to free companies from routine work, not to chase the model of the moment. Hy3 is a real step forward: it lowers computational cost, reduces hallucinations and gives control over data back to you. But technology alone doesn't create value. The real differentiator is the ability to integrate these tools into concrete, measurable and scalable workflows.

Picture a B2B manufacturing company that wants to automate the production of technical data sheets in Italian for the European market. The right model isn't necessarily the biggest or the most famous: it's the one that works for its specific use case, in its language, on its infrastructure. That's exactly the kind of evaluation we do with our clients: from choosing the model to designing the workflow, all the way to monitoring results.

The Italian AI market is accelerating, as the data from the Politecnico di Milano AI Observatory shows. The gap between large companies and SMEs is widening. Those who keep waiting risk starting from a position of structural disadvantage. The article on Meta Glimmer and Zuckerberg's AI vision also shows how quickly the open-weight landscape is being redrawn.

If you're evaluating how to integrate AI into your processes, you can explore our strategic AI integration service, the AI training for companies program or our data processing and business intelligence service. For a direct conversation, our team is available through the contact page.

Frequently Asked Questions

The right model isn't the biggest one. It's the one that solves your specific problem, in your language, on your infrastructure.