Tencent Releases and Open-Sources Tencent Hy4 Preview
Tencent has unveiled Hy4 preview, a 770B parameter, 1M+ token large language model, positioning it as a top-tier open-source competitor. This release is garnering attention on HN for its impressive real-world productivity claims across coding, office, and scientific research. The model's unique self-optimization capabilities, where it contributed to its own development and inference improvements, make it a noteworthy technical achievement.
The Lowdown
Tencent has officially launched and open-sourced its Hy4 preview, a significant advancement in large language models designed to tackle a wide array of real-world productivity tasks. This next-generation model boasts an impressive scale and a focus on practical applications, aiming to set a new standard for open-source AI performance.
- Model Specifications: Hy4 preview features 770 billion total parameters, 49 billion active parameters, and a context window exceeding 1 million tokens.
- Accessibility: It's available as an open-source model, through Tencent's products like WorkBuddy and CodeBuddy, and via API access through Tencent Cloud TokenHub and OpenRouter. Initial free access is offered on WorkBuddy and CodeBuddy for two weeks.
- Performance Benchmarks: Internal blind evaluations showed Hy4 preview slightly outperforming competitors like GLM-5.3 and Kimi K3, underscoring its top-tier status among open-source models.
- Productivity Focus: Developed with high-quality, expert-curated training data, Hy4 excels in:
- Software Engineering: Stronger understanding, planning, debugging, and validation for long-context tasks, enhancing front-end development.
- Office & Analytics: Improved understanding of complex environments, enhanced financial analysis, data analysis, and cross-document collaboration.
- Game Development: Ability to generate playable prototypes from natural language and integrate with game engines for multi-turn refinement.
- Scientific Research: Stronger capabilities in understanding, reasoning, and solving complex problems in AI, molecular dynamics, condensed-matter physics, and mathematics.
- Self-Improvement: Uniquely, Hy4 preview participated in its own development, optimizing training methods, data strategies, and evaluation frameworks, establishing an early-stage recursive self-improvement loop.
- Autonomous Optimization: The model also autonomously analyzed and optimized its inference system, leading to a 31.8% increase in end-to-end throughput.
- Cost Efficiency: API pricing is set at competitive rates (e.g., USD 0.834 per million input tokens).
Overall, Tencent Hy4 preview signals a major stride in accessible and powerful AI, emphasizing both high performance and a unique self-improving architecture, promising continuous evolution through real-world feedback via its 'preview-first' approach.