Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
Evolving story · 2 updatesTencent’s Hy3 MoE Model LaunchTimeline →Tencent’s AI lab released Hy3, an open-source MoE model with 295B total parameters but only 21B active per token, featuring a 256K context window and strong performance on SWE-Bench Verified.
- Hy3 is an open-source MoE model with 295B total parameters but only 21B active per token, optimizing efficiency.
- Features a 256K context window, enabling long-form reasoning and agentic tasks.
- Achieves 78.0 on SWE-Bench Verified with lower hallucination rates than prior models.
- Available under Apache 2.0 and free on OpenRouter until July 2026.
Tencent’s AI research team, Hy, has unveiled Hy3, a groundbreaking open-source Mixture-of-Experts (MoE) model that balances massive scale with efficiency. The model boasts 295 billion total parameters but activates just 21 billion per token, significantly reducing computational overhead while maintaining high performance. Hy3 introduces a 256K context window, enabling it to handle long-form reasoning, agentic tasks, and extended document analysis without losing coherence.
Performance benchmarks highlight Hy3’s capabilities, with a score of 78.0 on SWE-Bench Verified and reduced hallucination rates compared to prior models. This positions it as a strong contender for tasks requiring precision and extended context. The model is released under the Apache 2.0 license, making it freely accessible for research and commercial use. Additionally, Tencent is offering free access via OpenRouter until July 21, 2026, lowering the barrier for developers to experiment with its capabilities.
Source: Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context. Read the full piece at the source.
Provides an efficient, high-performance MoE model for research and deployment with minimal compute costs.
Enables cost-effective scaling of AI applications requiring long context and reasoning.
Signals Tencent’s growing influence in open-source AI and competitive positioning in MoE models.
Demonstrates advancements in balancing model size with practical efficiency.
- Mixture-of-Experts (MoE)
- A model architecture where only a subset of parameters (experts) are activated per input, improving efficiency without sacrificing performance.
- SWE-Bench Verified
- A benchmark evaluating AI models on software engineering tasks, measuring their ability to solve real-world coding challenges.
NVIDIA and Hugging Face Bring New Models and Frameworks to LeRobot for the Open Robotics Community
LeRobot v0.6.0: Imagine, Evaluate, Improve
OpenComputer | An Open Source Computer Built For Agents.
