Tencent open-sources Hy4 preview, an enterprise-first 770B model

Hy4 preview is a massive open-source Mixture of Experts model from Tencent Hunyuan, packing 770 billion total parameters with 49 billion active per token. Built with a 1M context window, it bypasses generic benchmark racing to focus strictly on real-world engineering, financial modeling, and scientific computing.

If you need a local model to tackle complex web development, inspect spreadsheets, or prototype games, this is built for that workflow.

What it can do

On real-world developer tasks, Hy4 preview is engineered for multi-step debugging, long-context project planning, and frontend aesthetics. On the Arena Web Dev leaderboard, it tied with GLM 5.3 Flash and outperformed Qwen 3.8 Max and DeepSeek V4 Pro.

Beyond coding, it handles multi-file enterprise finance analysis, direct playable game prototyping using game engines, and scientific tasks like molecular dynamics simulations. In Tencent's blind test involving 163 internal experts across 203 real-world engineering tasks, Hy4 preview scored 2.99 out of 4, edging out GLM 5.3 at 2.92 and Kimi K3 at 2.94.

Under the hood and hardware specs

Under the hood, the architecture adopts gated sparse attention, index caching, and identity hyperconnections to streamline information flow across layers, taking cues from recent designs like DeepSeek and GLM.

The standout innovation is recursive self-improvement. Hy4 preview actually helped develop itself by proposing training tweaks, running experiments, and optimizing low-level kernels based on performance logs. By identifying system bottlenecks autonomously, it boosted end-to-end inference throughput by 31.8%.

Running it locally requires serious iron. The unquantized model sits at roughly 1.56 TB. Fortunately, community contributors have already released GGUF versions, including a 1-bit quantization that brings the footprint down to 229 GB. The model is released under the permissive Apache 2 license.

If you do not have a server rack handy, it is accessible via Tencent Yuanbao, ima, and CodeBuddy or WorkBuddy, where it comes with two weeks of free access. API access is also available through Tencent Cloud Tokenhub and OpenRouter at competitive pricing.

Why it matters

For enterprise developers and research labs with high-end clusters, Hy4 preview gives you an Apache 2-licensed powerhouse tailored for codebases and technical documentation, without sending proprietary data to a third-party closed API.

For the broader ecosystem, Tencent is signaling that open-source frontier models are no longer limited to smaller Flash variants. More importantly, letting a frontier model design its own kernels and iterate its own training loop marks a real milestone for recursive model improvement.