A 295B-parameter MoE language model for agentic and productivity tasks.
Visit WebsiteFreeVisit Website
Tracked since2026
0 reviews trackedThe Bottom Line
Entry price
Free, no paid tier
Biggest pro
Outperforms models with 2-5x more total parameters on agentic and productivity tasks.
Biggest con
Requires high-memory GPUs (e.g., H20-3e) for deployment due to large total parameter count.
TL;DR - Tencent Hy3
- 295B-parameter MoE model with 21B active parameters, rivaling models 2-5x larger.
- Significant improvements in agentic capabilities and task productivity across coding, office work, and financial modeling.
- Enhanced reliability with reduced hallucination rate (from 12.5% to 5.4%) and better multi-turn conversation tracking.
Pricing: Free forever
Best for: Individuals & startups
What is Tencent Hy3?
Hy3 is a large-scale Mixture-of-Experts (MoE) language model developed by Tencent. It features 295 billion total parameters with only 21 billion activated per token, making it cost-effective for deployment. The model excels in agentic capabilities, reasoning, long-context understanding, and various productivity tasks such as coding, financial modeling, frontend design, and game development. It also emphasizes reliability with improvements in tool-call stability, anti-hallucination, and multi-turn intent tracking.
Pros & Cons
Pros
- Outperforms models with 2-5x more total parameters on agentic and productivity tasks.
- Reliability improvements in tool calls and output formats make it production-ready.
- Open-source weights available on multiple platforms, including Hugging Face and ModelScope.
Cons
- Requires high-memory GPUs (e.g., H20-3e) for deployment due to large total parameter count.
- FP8 quantized model available, but it still demands substantial hardware resources.
Preview
Key Features
Mixture-of-Experts (MoE) architecture with 192 experts and 8 activated per token.Supports 256K context length and BF16 precision.Stable tool call and output format consistency across multiple agent scaffoldings.Fine-grained anti-hallucination training to reduce factual errors and logical contradictions.Joint optimization of supervised fine-tuning and reinforcement learning for better multi-turn intent tracking.
Pricing
Free
Tencent Hy3 is completely free to use with no hidden costs.
Reviews
$99Free with your review
Write a reviewReview Tencent Hy3, get a free AI guide
Share your experience and we will send you Improve Your Thinking Patterns Using ChatGPT, free.
Best Tencent Hy3 Alternatives
Top alternatives based on features, pricing, and user needs.
Claude CodeFreemium
Agentic coding tool that lives in your terminal
KiteFree
An AI-powered coding assistant designed to help developers write code faster.
GitHub CopilotPaid
AI pair programmer that suggests code
SafuraiFreemium
Your AI coding assistant for faster, smarter, and more efficient development.
Still deciding?
Most buyers shortlist 2 or 3 tools before committing. Pull a side-by-side comparison or browse the full alternatives shortlist below.
Explore More
Tencent Hy3 FAQ
How does Hy3 handle long-context conversations without losing track of intent?
Hy3 uses a joint optimization of supervised fine-tuning and reinforcement learning to improve coreference resolution, ellipsis recovery, and multi-turn constraint inheritance. This reduces multi-turn issue rates from 17.4% to 7.9% in internal tests.
What agent scaffoldings are compatible with Hy3, and how consistent are the results?
Hy3 generalizes across different agent scaffoldings. On SWE-Bench Verified, accuracy variance across scaffoldings like CodeBuddy, Cline, and KiloCode remains within 4%.
How was Hy3 evaluated beyond public benchmarks?
Tencent conducted a blind evaluation with 270 experts using real work tasks. Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4, with the strongest advantages in frontend development, data & storage, and CI/CD tasks.
What hardware is recommended for serving Hy3?
Due to its 295B total parameters, Hy3 requires high-memory GPUs like the H20-3e for serving on 8 GPUs. Production deployment is supported via vLLM or SGLang.
Is there a quantized version of Hy3 available?
Yes, Hy3-FP8 is an FP8 quantized instruct model that is also open-sourced on platforms like Hugging Face, ModelScope, GitCode, and CNB.
What context window size does Hy3 support?
Hy3 supports a context length of 256K tokens, enabling it to handle very long documents or extended multi-turn conversations.
How does Hy3's reasoning mode work for different task complexities?
Hy3 offers a reasoning_effort parameter: 'no_think' for direct responses, 'low' for moderate reasoning, and 'high' for deep chain-of-thought reasoning. It is recommended to set 'high' for complex tasks like math and coding.
What specific productivity tasks does Hy3 excel at?
Hy3 has shown strong performance in coding, office work, financial modeling, frontend design, and game development, making it suitable for a wide range of professional applications.
Source: huggingface.co