Tencent has officially released and open-sourced Tencent Hy4 Preview, a next-generation large language model built on a 770-billion-parameter mixture-of-experts architecture with 49 billion active parameters per pass. The model ships with a context window exceeding one million tokens and is positioned as a top-tier open-source system aimed squarely at real-world productivity, spanning coding, office workflows, game development, and scientific research. Tencent says the jump in capability comes from a combination of larger scale, longer context, more data, and major advances in both pre-training and post-training.
What Tencent Hy4 Preview Brings to the Table
The headline numbers are striking. Tencent Hy4 Preview is a 770B-parameter MoE model with 49B active parameters, and it can keep more than one million tokens of context in mind at once. That scale, combined with Tencent’s focus on high-quality, expert-curated training data drawn from software engineering, gaming, finance, and security, is what the company credits for the model’s productivity gains.
Tencent also emphasizes deep co-design with its products. Hunyuan worked alongside teams building CodeBuddy and WorkBuddy so that improvements in the model translate directly into the tools end users touch every day. The result is a model designed less for benchmark theater and more for shipping useful work in long, messy, multi-document tasks.
Blind Evaluation Results: Beating GLM-5.3 and Kimi K3
Tencent ran an internal blind evaluation with 163 experts grading 203 engineering tasks. Tencent Hy4 Preview scored an average of 2.99 out of 4.00, narrowly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. While the margins are slim, the win is meaningful given how directly the contenders were compared on real productivity work rather than synthetic benchmarks.
For software engineering specifically, Tencent Hy4 Preview shows stronger understanding, planning, debugging, and validation on long-context development tasks, along with improved visual quality and interaction in front-end work. In office scenarios, the model handles complex working environments, financial analysis, and cross-document collaboration, supporting full workflows from information processing to the creation of documents, spreadsheets, and presentations.
Game Development and Scientific Research
Game developers get a notable upgrade. Tencent Hy4 Preview can generate a playable prototype from a single natural-language request and work effectively with game engines, allowing developers to refine complex projects through multi-turn interactions. This pushes prototyping significantly earlier in the development cycle, where iteration speed tends to matter most.
On the research side, the model shows stronger capabilities in understanding, reasoning through, and solving complex problems across AI research and development, molecular dynamics simulation, condensed-matter physics, and fundamental mathematics. These gains position Tencent Hy4 Preview as a useful assistant not just for productivity suites, but also for laboratory and engineering teams working on technically demanding problems.
Self-Improving AI: The Model Helped Build Itself
One of the most unusual aspects of this release is the model’s role in its own development. Tencent Hy4 Preview participated for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators. It proposed approaches, ran experiments, iterated based on results, and fed its code, logs, and feedback into subsequent rounds of exploration. Tencent describes this as an early-stage recursive self-improvement loop.
The same pattern extended to infrastructure. The model autonomously analyzed bottlenecks in its inference system and ran multiple rounds of optimization, including operator fusion and communication tuning. Those changes lifted end-to-end throughput by 31.8% compared with the baseline, with consistent gains across different context lengths and concurrency levels. Few shipping models today contribute meaningfully to their own serving stack.
Pricing, Availability, and Access
Tencent Hy4 Preview is available as an open-source model and through WorkBuddy, CodeBuddy, Yuanbao, ima, and other Tencent products. Enterprise developers can also connect via API through Tencent Cloud TokenHub and OpenRouter. WorkBuddy and CodeBuddy users get free access to the preview for two weeks, and free access to the previous-generation Hy3 has been extended through September 30.
API pricing is set at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits, keeping Tencent Hy4 Preview competitive with other frontier-tier offerings while remaining accessible to a wider range of teams.
What Comes Next for the Hy4 Series
Tencent says it follows a preview-first release strategy, using real-world feedback to inform each subsequent official launch. The next batch of models in the Hy4 series is expected to roll out soon, suggesting that what we see today is the opening move in a longer cadence of releases rather than a one-off drop.
With its combination of long context, expert-graded productivity wins over GLM-5.3 and Kimi K3, competitive API pricing, and a notable recursive self-improvement story, Tencent Hy4 Preview marks one of the more ambitious open-source LLM launches of the year. Teams evaluating new foundation models for coding, office, or research workloads now have another serious option to put through its paces, and Tencent Hy4 Preview is widely available to do exactly that.
Source: https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/