Tencent Hy4 Preview: The 770B-Parameter Open-Source Giant Built for Extreme Productivity

The artificial intelligence landscape is shifting once again. Released on August 28, 2026, Tencent officially pulled back the curtain on the Hy4 Preview (Hunyuan-4). Claiming a definitive spot at the open-source frontier, Hy4 is not just another conversational chatbot—it is a heavy-duty engine designed to solve complex, long-horizon tasks that make lesser models stumble.

Whether you are a software engineer, an enterprise workflow architect, or a developer looking for the best AI API unit economics on the market, Tencent’s latest release demands your attention. In this deep dive, we will break down the architecture, the real-world benchmarks, and why Hy4's pricing structure could change how you build AI agents.

Under the Hood: Massive Scale Meets MoE Efficiency

Scaling an AI model effectively requires a delicate balance between sheer power and inference speed. Tencent scaled the Hy4 preview on three critical fronts: model size, context length, and training data quality.

• Total Parameters: 770 Billion
• Active Parameters: 49 Billion (Mixture-of-Experts architecture ensures fast, cost-effective inference)
• Context Window: 1 Million Tokens (capable of ingesting massive codebases, entire books, and years of financial data in a single prompt)

By leveraging a stronger pre-training pipeline and a substantially larger post-training run, Tencent achieved the largest generation-over-generation capability gain they have ever measured.

The Four Pillars of Productivity: Where Hy4 Dominates

Instead of relying purely on generic internet scrapes, Tencent partnered with internal experts—game developers, security analysts, and software engineers—to build training data based on real work shipped to production. The result is a model purpose-built for four key domains:

1. Long-Horizon Software Engineering

Hy4 excels at understanding, planning, and debugging complex development tasks. It also demonstrates a remarkable leap in "visual taste" for front-end work. For example, when tasked to build a 3D desktop viewer for historic architecture, Hy4 didn't just write functional Three.js code; it successfully adhered to strict design constraints—implementing a fixed three-column layout, rice-paper backgrounds, serif typography, and drag-and-rotate mechanics while actively avoiding generic "dark tech" aesthetics.

2. Enterprise Office and Data Analysis

For financial analysts and project managers, Hy4 is a massive time-saver. It can take messy context spread across dozens of disparate files and convert them into polished, shareable artifacts. In a practical test, Hy4 was able to review 24 separate reimbursement claims by cross-referencing multiple policy versions, budgets, and email ledgers. It output a clean summary table with cited reasoning, verified arithmetic, and flagged suspected fraud.

3. Game Development and Prototyping

Game developers can turn a single prompt into a playable prototype. Hy4 integrates seamlessly with engines like Unity and Tuanjie. In one instance, the model generated a complete third-person 3D game featuring a penguin exploring a snowfield, complete with inventory systems, animation blend trees, combat mechanics, and custom HUDs—with all variables neatly exposed as editor parameters for the developer to tweak.

4. Advanced Scientific Research

Moving beyond code and text, Hy4 shows solid reasoning in molecular dynamics, condensed matter physics, and pure mathematics. Its most impressive feat? Agentic orchestration. Hy4 successfully managed several smaller Codex sessions in parallel, acting as the lead researcher. It optimized multiple evaluation targets at once and adjusted its research direction dynamically as new results came in, proving its ability to organize complex R&D workloads.

Benchmarks: Outperforming the Competition

To ensure these capabilities translate to real-world value, Tencent ran a blind, side-by-side evaluation. They tasked 163 internal experts to rate model outputs across 203 grueling engineering tasks. The results prove Hy4 is a top-tier contender:

• Hy4 Preview vs. GLM 5.3: Hy4 came out ahead with an average score of 2.99 vs 2.92 (46.8% wins, 12.8% ties, 40.4% losses).
• Hy4 Preview vs. Kimi K3: Hy4 won again with an average score of 2.99 vs 2.94 (51.2% wins, 7.9% ties, 40.9% losses).

These gains are already being felt inside Tencent, as Hy4 actively powers internal productivity tools like CodeBuddy, WorkBuddy, Yuanbao, and ima.

API Pricing: Incredible Unit Economics for Developers

For developers and AI product builders, the API pricing is arguably the most exciting part of this announcement. Tencent has committed to keeping this frontier model affordable and accessible. Currently available on Tencent Cloud and OpenRouter, the pricing (per 1 Million tokens) is aggressively competitive:

• Cached Input: $0.042
• Standard Input: $0.834
• Output: $2.501

The Developer Advantage: The cached input price of just $0.042 per million tokens is an absolute game-changer. If you are building Retrieval-Augmented Generation (RAG) applications, massive document analyzers, or conversational agents that rely on heavy system prompts, this pricing structure allows you to scale your application with incredible profit margins.

Known Issues and What's Next

Tencent is transparent that this is an early preview version of Hy4. While highly capable, it currently ships with a few known quirks. Most notably, the model sometimes spends longer than necessary reasoning through complex tasks and has a tendency to "over-verify" its own work.

However, Tencent’s strategy of shipping early and iterating based on real-world friction—the exact strategy that made Hy3 a success—means we can expect rapid optimizations in the coming weeks. There is still massive headroom left in both the pre-training and post-training phases.

Final Thoughts: Should You Try Hy4?

Absolutely. The Tencent Hy4 Preview represents a massive leap forward for the open-source community. By combining a 770B-parameter intelligence with an ultra-efficient 49B active MoE architecture and a 1M token context window, it brings enterprise-grade processing to the masses.

Whether you want to automate complex office workflows, bootstrap game prototypes, or build high-margin AI SaaS apps taking advantage of the ultra-cheap cached input pricing, Hy4 provides the tools to get it done.

Ready to start building? You can access the Hy4 Preview API today via Tencent Cloud or OpenRouter. Integrate it into your workflows and experience the new standard in open-source AI.