Tencent Unveils Hy3: An Open 295B MoE Model for Advanced AI Workflows

Introducing Hy3: Tencent’s Latest Open-Source AI Breakthrough

At a glance, Tencent’s Hy team has officially launched Hy3, a groundbreaking Mixture-of-Experts (MoE) model. This powerful new entry into the open-source AI landscape boasts an impressive 295 billion total parameters, yet intelligently activates only 21 billion parameters per token, striking a balance between immense capability and computational efficiency. Released under the Apache License 2.0, Hy3 is specifically engineered to excel in complex reasoning tasks, sophisticated agentic workflows, and handling exceptionally long context windows, up to 256,000 tokens.

Meanwhile, This release marks a significant step forward, offering developers and researchers a robust tool designed not just for raw performance, but also with a strong emphasis on production reliability and practical application across various domains.

Understanding Hy3’s Innovative Architecture

At its core, Hy3 leverages a sparse Mixture-of-Experts (MoE) architecture. This design is crucial for its efficiency:

• Expert Network: Hy3 incorporates 192 distinct experts.
• Top-8 Routing: For any given token, only 8 of these experts are activated, significantly reducing the computational load compared to dense models of similar total parameter count.

In practical terms, Beyond its MoE foundation, Hy3 integrates another key innovation: a Multi-Token Prediction (MTP) layer. This feature is designed to accelerate decoding by predicting several tokens simultaneously, leading to faster inference. MTP is readily supported through speculative decoding mechanisms in popular serving frameworks like vLLM and SGLang.

Key Architectural Specifications

• Architecture: Mixture-of-Experts (MoE)
• Total Parameters: 295 Billion
• Activated Parameters: 21 Billion (per token)
• MTP Layer Parameters: 3.8 Billion
• Context Length: 256,000 tokens
• Experts: 192 (top-8 activated)
• Supported Precisions: BF16, with a separate Hy3-FP8 checkpoint available to further reduce memory footprint for more cost-effective serving.

Benchmark and Real-World Performance

Tencent’s research team has rigorously benchmarked Hy3 across a spectrum of tasks, showcasing its capabilities in coding, agentic scenarios, and STEM fields.

• Coding Prowess: Hy3 achieved strong scores on coding benchmarks, including 78.0 on SWE-Bench Verified, 57.9 on SWE-Bench Pro, and 75.8 on SWE-Bench Multilingual. It also scored 71.7 on Terminal-Bench 2.1 and 28.0 on DeepSWE.
• STEM and Reasoning Excellence: The model truly shines in demanding STEM and reasoning challenges, reporting 90.4 on GPQA Diamond, 72.0 on USAMO 2026, and 90.0 on IMOAnswerBench.

For example, In a compelling blind test involving 270 experts and 312 valid comparisons on real-world workflows, Hy3 outperformed GLM-5.1, scoring 2.67 out of 4 compared to GLM-5.1’s 2.51. This edge was particularly evident in practical applications like frontend development, CI/CD, and data/storage tasks.

Built for Reliability: Addressing Production Challenges

A core focus of the Hy3 release is its enhanced production reliability. The research team specifically addressed three critical failure modes:

• Tool Calling and Output Formatting: Significant improvements have been made to baseline stability, preventing issues like invalid calls that lead to infinite loops. Hy3 now generalizes well across various agent scaffoldings, with accuracy variance on SWE-Bench Verified remaining within 4% across different frameworks.
• World Knowledge and Anti-Hallucination: Hy3 is trained to provide grounded answers and flag instances where evidence is missing. Internal evaluations show a substantial reduction in hallucination rates from 12.5% to 5.4%, and commonsense error rates from 25.4% to 12.7%.
• Multi-Turn Intent Tracking: Through joint SFT and RL, improvements in coreference and constraint tracking have dramatically reduced internal issue rates from 17.4% to 7.9%. On the MRCR long-dialogue benchmark, scores climbed from 42.9% to 75.1%.

How to Access and Use Hy3

That said, Hy3 offers an OpenAI-compatible API, making it easy to integrate into existing workflows. Deployment is streamlined using vLLM or SGLang, after which you can call the endpoint. A unique feature is the reasoning_effort flag, which allows users to control the model’s ‘thinking’ process:

• no_think (default): For direct, straightforward answers.
• low: For moderately complex tasks.
• high: For demanding tasks like advanced math, intricate coding, or multi-step problem-solving, enabling a deep chain-of-thought process.

Tencent recommends using a temperature of 0.9 and top_p of 1.0 for optimal results. For those without local hardware, Hy3 can be tested via OpenRouter’s tencent/hy3:free route, though this free tier is scheduled to end on July 21, 2026.

Practical Applications: Where Hy3 Shines

Interestingly, Hy3 is specifically optimized for agent-style and long-context applications, opening up a variety of powerful use cases:

• Coding Agents: Feed an entire code repository into its 256K context window and instruct Hy3 to fix failing tests with reasoning_effort=”high”. Its stable tool calls facilitate multi-file edits.
• Document Processing: Analyze lengthy contracts or legal filings. The anti-hallucination training helps prevent fabricated clauses or misquotes.
• Financial Analysis: Process complex financial reports, combining tables and prose within a single prompt. Request grounded summaries that accurately flag missing data instead of making assumptions.
• Frontend and Game Development: Generate components (e.g., React) or small game loops. The blind test results highlighted Hy3’s advantage over GLM-5.1 in frontend development tasks.

Hy3 vs. The Competition: A Strategic Choice

While Hy3 is a formidable model, it’s also important to understand its positioning against other large language models. When compared to GLM-5.2, which is a roughly 744B MoE model with about 40B active parameters, Hy3 offers a different value proposition.

However, Hy3, at less than half the total size (295B) and with a significantly smaller active footprint (21B vs. ~40B), trades some raw coding accuracy for greater resource efficiency.

For instance, on SWE-Bench Verified, GLM-5.2 scores 84.2 compared to Hy3’s 78.0. However, this smaller active footprint translates into lower GPU costs and easier self-hosting, making Hy3 an attractive option for organizations mindful of operational expenses and infrastructure.

Deployment Considerations

Given Hy3’s 295 billion total parameters, serving it requires substantial memory. Tencent’s research team recommends deploying Hy3 on hardware with ample GPU memory, such as 8 H20-3e GPUs or similar high-capacity cards. Both vLLM and SGLang provide ready-to-use recipes with MTP enabled for seamless deployment.

Meanwhile, For those looking to optimize deployment further, Tencent offers its AngelSlim toolkit, which encompasses quantization, low-bit methods, and speculative sampling techniques. Additionally, a complete finetuning pipeline for Hy3 is available, allowing for custom model adaptations.

Expert Perspective

A practical read on Tencent Hy3 MoE model starts with experts. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make Tencent Hy3 MoE model a meaningful reference point across tencent.

For decision-makers, the useful lens is not the headline alone but how billion changes priorities once organizations have to respond.

Frequently Asked Questions

Why is Tencent Hy3 MoE model important?

Introducing Hy3: Tencent’s Latest Open-Source AI Breakthrough At a glance, Tencent’s Hy team has officially launched Hy3, a groundbreaking Mixture-of-Experts (MoE) model.

What impact could Tencent Hy3 MoE model have?

This powerful new entry into the open-source AI landscape boasts an impressive 295 billion total parameters, yet intelligently activates only 21 billion parameters per token, striking a balance between immense capability and computational efficiency.

What should readers watch next with Tencent Hy3 MoE model?

Released under the Apache License 2.0, Hy3 is specifically engineered to excel in complex reasoning tasks, sophisticated agentic workflows, and handling exceptionally long context windows, up to 256,000 tokens.

How does this relate to experts?

It connects because the article frames experts as one of the clearest areas where the topic may be felt in practice.

Conclusion

What matters next is how the immediate response turns into lasting change. Tencent’s Hy3 represents a compelling new option in the open-source MoE landscape. By combining a massive parameter count with an efficient active parameter design, advanced features like Multi-Token Prediction, and a strong focus on production reliability, Hy3 is well-positioned to empower developers and enterprises tackling demanding AI tasks. Its strategic balance between performance and resource efficiency makes it a notable contender for next-generation agentic and long-context applications.

Source: https://www.marktechpost.com/2026/07/06/tencent-releases-hy3-open-295b-moe-model/