Skip to main content

Tencent's Hy3preview AI Model Breaks New Ground in Practical Intelligence

Tencent's AI Leap: Hy3preview Goes Open-Source

Marking a major milestone in artificial intelligence development, Tencent has released Hy3preview - the first model trained after reconstructing its Huan Yuan infrastructure. This isn't just another language model; it's a practical problem-solver designed to think both fast and deep.

Image

What Makes Hy3preview Special?

With 295 billion total parameters (21 billion active), Hy3preview supports context lengths up to 256K tokens. But numbers alone don't tell the story. Tencent rebuilt their entire approach around three principles:

  • Systematic capabilities that work together like a well-oiled machine
  • Real-world evaluation beyond standard benchmarks
  • Cost-effectiveness that makes advanced AI accessible

"This is just our first step," says Yao Shunyu, Tencent's Chief AI Scientist. "We're opening this model to gather genuine feedback before the official Hy3 launch."

Brainpower Where It Counts

The model shines in areas that matter:

1. Context Mastery Hy3preview digests messy, lengthy information with ease - crucial for real-world applications. Tencent developed special tests (CL-bench) to prove its mettle.

Image

2. Reasoning Champion It aced Tsinghua University's math doctoral exam and national biology competitions, showing it can tackle complex problems beyond simple Q&A.

3. Coding & Agent Skills Developers will appreciate its improved performance on coding tasks (54% faster first response) and agent workflows (495-step stability). The numbers speak for themselves:

Image Image Image Image

Already Powering Your Favorite Apps

The proof is in the pudding - Hy3preview is already enhancing experiences across Tencent's ecosystem:

  • Yua: More human-like conversations with better emotional intelligence
  • Peace Elite: NPCs that think and react like real players
  • QQ Browser: Smarter searches with better comprehension
  • Tencent Docs: AI-generated presentations with 20% faster creation

The improvements aren't just technical - they're making everyday interactions smoother and more natural.

Affordable Intelligence

The best part? This power comes at consumer-friendly prices on Tencent Cloud's TokenHub platform:

  • Input: ¥1.2/million tokens
  • Output: ¥4/million tokens
  • Personal plan: Just ¥28/month

Image Image

Key Points:

  • Hybrid Expert Model: Combines fast/slow thinking approaches
  • Practical Focus: Designed for real-world business applications
  • 40% Efficiency Boost: Faster responses at lower cost
  • Wide Adoption: Already integrated across Tencent products
  • Developer Friendly: Affordable pricing and open-source availability

Enjoyed this article?

Subscribe to our newsletter for the latest AI news, product reviews, and project recommendations delivered to your inbox weekly.

Weekly digestFree foreverUnsubscribe anytime

Related Articles

DeepSeek V4 Launches with Two Versions: What Developers Need to Know
News

DeepSeek V4 Launches with Two Versions: What Developers Need to Know

DeepSeek has unveiled its latest AI model, V4, offering two distinct versions tailored for different needs. The Flash version targets cost-effective, high-speed applications, while the Pro version handles complex reasoning tasks. With transparent pricing and a focus on caching optimization, this release could reshape how businesses approach AI integration. Here's a breakdown of the key features and what they mean for developers.

April 24, 2026
AI DevelopmentMachine LearningTech Innovation
Tencent's Hunyuan 3.0 AI Model Leaps Forward in Coding Prowess
News

Tencent's Hunyuan 3.0 AI Model Leaps Forward in Coding Prowess

Tencent has unveiled its latest AI powerhouse, Hunyuan 3.0, showcasing remarkable improvements in programming capabilities. With test scores jumping from 53% to 74.4%, this new model demonstrates a 40% performance boost over its predecessor. The tech giant's strategic hiring of AI expert Yao Shunyu appears to be paying off as they position themselves against industry leaders like OpenAI and DeepSeek.

April 23, 2026
Artificial IntelligenceTech InnovationProgramming Tools
Singapore Emerges as AI's Neutral Ground Amid US-China Tensions
News

Singapore Emerges as AI's Neutral Ground Amid US-China Tensions

As tech competition heats up globally, Singapore is quietly transforming into a strategic hub for AI companies seeking to avoid geopolitical pitfalls. Both Chinese startups and American tech giants are flocking to this Southeast Asian city-state, drawn by its business-friendly policies, efficient visa processes, and legal protections. While offering refuge from political tensions, Singapore's growing AI prominence also raises questions about the future of technological sovereignty.

April 24, 2026
Artificial IntelligenceSingapore TechGeopolitics
AI Startup DeepSeek Eyes $20B Valuation in Talent War with Tech Giants
News

AI Startup DeepSeek Eyes $20B Valuation in Talent War with Tech Giants

DeepSeek, the self-funded AI research lab, is seeking its first external investment at a staggering $20 billion valuation. The move comes as tech titans Tencent and Alibaba circle the promising startup, poaching key researchers. This financing round - while relatively small in dollar terms - serves primarily to boost employee stock values and retain top talent in the fiercely competitive AI market.

April 24, 2026
Artificial IntelligenceTech StartupsChina Tech
DeepSeek-V4 Arrives: AI That Remembers Like Never Before
News

DeepSeek-V4 Arrives: AI That Remembers Like Never Before

DeepSeek has unveiled its latest AI model, the V4 series, featuring breakthrough capabilities in handling long documents and complex reasoning. The new models come in two flavors - a powerhouse Pro version and a budget-friendly Flash option - both capable of processing up to one million words at once. What makes this release special? It brings professional-grade AI within reach of everyday users while setting new standards for open-source intelligence.

April 24, 2026
AI InnovationDeepSeekOpen Source AI
Meet the 13-Person Team Behind GPT Image2's AI Art Revolution
News

Meet the 13-Person Team Behind GPT Image2's AI Art Revolution

The surprisingly small team behind GPT Image2 has achieved what many thought impossible - completely redesigning AI image generation in just four months. Led by former Google researcher Chen Boyuan, this tight-knit group has created what they call 'GPT for images,' solving persistent problems like text rendering and spatial understanding. Their work demonstrates how focused innovation can outperform massive corporate teams.

April 23, 2026
AI ArtMachine LearningTech Innovation