On August 28, 2026, Chinese IT giant Tencent's Hunyuan announced the release and open-sourcing of its next-generation large language model Hy4 preview. The model has 770B total parameters and 49B activated parameters, supports a context window of over 1 million tokens, and is positioned as the most capable model in the Hunyuan series. The company emphasized significant progress in hard-to-deploy scenarios such as long-duration software engineering, document-intensive office work, and scientific research, and said the model has entered the top tier of open-source models.

Details: scale expansion and productivity orientation

Hy4 preview expands model scale, context length, and training data in parallel. Tencent said stronger pretraining and substantially extended post-training together produced the largest generational capability jump. Architecturally it adopts a mixture-of-experts (MoE) design: 770B total parameters with about 49B activated per token; a 78-layer backbone with most layers as MoE (each layer contains 256 routing experts plus one shared expert; each token activates the top-8 routing experts plus the shared expert), and it supports a 1M token context.

The model is optimized for real productivity scenarios:

Software engineering: improved understanding, planning, debugging and verification for long-cycle development tasks, with further enhancements in frontend visual and interaction quality.
Office and analytics: capable of handling complex cross-file context and turning it into deliverables such as documents, spreadsheets and presentations, with higher accuracy in data analysis, equations and financial models.
Game development: able to generate playable prototypes from single prompts and collaborate smoothly with game engines, supporting multi-turn iteration.
Scientific research: shows stronger understanding and reasoning on hard problems in AI research, molecular dynamics, condensed matter physics and pure mathematics. The company noted the model can manage multiple Codex sessions in parallel and dynamically adjust research directions, acting as a "researcher" to coordinate optimization in small-model post-training tasks, surpassing isolated Codex exploration.

Tencent organized 163 internal experts to conduct blind tests on 203 real engineering tasks: Hy4 preview achieved an average score of 2.99/4, slightly above GLM-5.3 (2.92) and Kimi K3 (2.94). The model is already available via Tencent Cloud TokenHub and OpenRouter, and has been integrated into products such as CodeBuddy, WorkBuddy, Yuanbao and ima. API pricing is: cached input $0.042 per 1M tokens, input $0.834 per 1M tokens, output $2.501 per 1M tokens. WorkBuddy and CodeBuddy offer two weeks of free use.

The company acknowledged this is an early release and that issues remain, such as overly long reasoning paths and excessive self-verification, and stated it will continue rapid iteration.

Technical and market background

Since rebuilding its AI infrastructure earlier in 2026, the Hunyuan model iteration pace has accelerated (Hy3 was released in July and Hy4 preview followed in under two months). Hy4 preview significantly increases parameter scale and context capability over its predecessor and emphasizes deep co-creation with Tencent internal products and experts so that model capabilities directly serve real workflows. The open-source strategy continues Hunyuan's positioning of being "open and affordable," aiming to benefit more developers and enterprises.

Potential impact: an upgrade in the competition for open-source productivity models

1. Impact on the open-source ecosystem

The combination of 770B total parameters, 49B activation and a 1M token context gives Hy4 preview strong competitiveness on lo