Tencent open-sources Hy4 preview with 1M context for coding agents

Tencent has released Hy4 preview as an open-weight Mixture-of-Experts model with 770 billion total parameters, 49 billion activated per token, and a context window exceeding one million tokens. The model is available through Tencent’s WorkBuddy and CodeBuddy products, Tencent Cloud TokenHub, and OpenRouter.

Hy4 preview targets long production tasks

Tencent positions Hy4 preview for software engineering, office work, data analysis, game development, and scientific research. The model repository describes a 78-layer architecture with Gated DeepSeek Sparse Attention and a native multi-token prediction layer for speculative decoding. Those details matter for operators evaluating whether the model can sustain long tool-using workflows rather than only answer isolated prompts.

The 1M-token context window is the headline capability. It gives coding agents room to retain large repositories, long issue threads, and generated artifacts in one working context, although actual throughput and quality will depend on the serving configuration and workload.

Access, pricing, and the free window

Hy4 preview is available globally in WorkBuddy and CodeBuddy, alongside Tencent products including Yuanbao and ima. Tencent says WorkBuddy and CodeBuddy will provide free access for two weeks after launch. The release does not define that window as unlimited, so users should treat it as a time-limited access offer rather than a capacity guarantee.

OpenRouter lists the model as tencent/hy4-preview at $0.834 per million input tokens and $2.501 per million output tokens, with cache-read pricing listed separately. Provider uptime and latency data are still immature because the model is new. Teams should start with bounded evaluations and watch output-token consumption before moving long-running agents into production.

What the reported evaluation does—and does not—show

Tencent reports an internal blind evaluation in which 163 experts rated 203 engineering tasks. Hy4 preview scored 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. TechNode independently reports those figures, but the benchmark remains Tencent-designed and Tencent-run; there is no independent benchmark confirmation in the release materials reviewed for this article.

Tencent also says Hy4 preview helped optimize parts of its own training and inference stack, producing a 31.8% end-to-end throughput improvement against a baseline. That is a vendor-reported systems result, not a general claim that the model autonomously improves arbitrary deployments.

The practical choice for developers

Hy4 preview changes the open-model shortlist for teams that need a large context window, tool use, and a model that can be reached through both hosted products and an API aggregator. The official repository provides the deployment path for teams that want to inspect or run the released artifacts, while OpenRouter offers a faster hosted test route.

The sensible next step is a task-level comparison against the models already in production: measure successful edits, tool-call reliability, latency, maximum useful context, and total output-token cost. Hy4 preview is released and accessible now, but its internal quality claims still need independent workload testing.

Sources and methodology

This report uses Tencent’s announcement as the primary source, then checks the official model repository, OpenRouter’s live model page, and independent TechNode coverage. Tencent’s internal benchmark and throughput claims are labeled as such; no independent validation is implied.