Tencent's Hy4 Preview: 770 Billion Parameters, Apache License, Open Weights

Tencent has released Hy4 preview, a new open-weights model built on a mixture-of-experts architecture and aimed squarely at high-complexity workloads: software engineering, research, and financial analysis. The weights are live on Hugging Face under an Apache-2.0 license—meaning any builder or enterprise can download, fine-tune, and deploy commercially without negotiating a proprietary agreement.

The scale and the licensing together are the story. This is not a research artifact locked behind an API.

Architecture: Scale With Efficiency

Hy4 preview carries 770 billion total parameters, but the mixture-of-experts design means only 49 billion parameters are active per request. That distinction matters operationally. Full-dense models at this parameter count demand hardware most teams can't provision; MoE architectures route each token through a subset of specialized experts, keeping active compute closer to what mid-tier clusters can absorb while preserving the reasoning capacity associated with much larger models.

The tradeoff is real—MoE models carry heavier memory footprints than their active-parameter count implies, because all expert weights must reside in memory even if only a fraction fires on any given forward pass. But for the target workloads Tencent named—long code reviews, document-heavy financial analysis, extended research synthesis—the architectural bet is defensible. These tasks reward depth of reasoning over raw throughput.

Context Window: One Million Tokens

Hy4 preview ships with a 1 million–token context window. That number has become a competitive signal in its own right, but it maps directly onto the use cases Tencent is pitching. A million tokens is roughly 750,000 words—enough to ingest an entire codebase, a multi-year financial filing history, or a corpus of research papers in a single pass.

For software engineering tasks specifically, long context changes what the model can actually do. Debugging across a sprawling repository, tracing a call stack through dozens of interdependent files, or generating code that must remain consistent with thousands of lines of existing logic—all of these become tractable when the model holds the full context rather than a summarized proxy of it. Enterprise and research users building internal tools around Hy4 will likely find the context ceiling is less of a bottleneck than their inference infrastructure.

Positioning: Hunyuan's Enterprise Tier

Tencent positioned Hy4 as part of its Hunyuan model family, which has targeted enterprise and research users throughout its development. The preview designation signals this is not a finished, production-hardened release—benchmarks, safety evaluations, and optimized inference paths are presumably still in progress. But publishing the weights now, before a full release, is a deliberate move: it lets the research community and early enterprise adopters begin integration work, surface edge cases, and generate feedback before Tencent finalizes the model.

The Apache-2.0 license amplifies this strategy. Restrictive licenses on models this large tend to fragment the ecosystem—teams build around them cautiously, if at all, because commercial deployment risk is unclear. Apache-2.0 removes that friction entirely. Startups, enterprises, and academic labs can all build on Hy4 preview without legal ambiguity.

This is the same licensing move that accelerated adoption of other open-weights releases, and Tencent appears to be making it deliberately rather than incidentally.

The Bigger Shift

Hy4 preview is another data point in the restructuring of who can field frontier-scale AI infrastructure. A 770-billion-parameter model with a permissive license and a million-token context window, available for direct download—that combination was not a realistic option for most builders two years ago. It is now.

The competitive pressure this creates runs in multiple directions: against proprietary API providers who charge for access to comparable capability, against other open-weights labs racing to hit similar parameter counts and context lengths, and against the assumption that Chinese AI labs are primarily building for domestic deployment. Tencent publishing Hy4 preview on Hugging Face with an Apache license is an explicitly global distribution move.

What gets built on top of it—and how quickly—will say more about the model's real impact than any benchmark Tencent releases alongside it.