Tencent's Hunyuan unit open-sourced Hy4 preview, a mixture-of-experts model with 770 billion total parameters, 49 billion activated per token, and a context window exceeding one million tokens, distributed through its WorkBuddy, CodeBuddy, Yuanbao, and ima products with API access.
The sparse-activation design is the real story. Firing only 49B of 770B parameters per token means Hy4 chases frontier-scale capability while keeping inference cost closer to a mid-sized dense model. That economics-first framing is exactly how China's open-weight camp — DeepSeek, Alibaba's Qwen, and now Tencent — is trying to reset the market: not by winning a single benchmark, but by collapsing the cost of 'good enough' reasoning and long-context work. When a permissively distributed model can ingest a million tokens, whole categories of retrieval-augmented plumbing become optional, and the moat of closed API providers narrows to safety, tooling, and enterprise trust rather than raw capability.
The strategic subtext is distribution. By bundling Hy4 into coding and productivity assistants while releasing weights, Tencent buys developer mindshare and turns its own product suite into a reference deployment. This is the same flywheel Meta ran with Llama, now with a Chinese center of gravity — which is precisely what makes it a governance question, not just a technical one, for buyers outside China.
For Japanese enterprises and SIers, this creates a genuine fork. On cost, open Chinese weights are increasingly hard to ignore for internal document analysis, code assistance, and long-context summarization workloads where running models on-premise or in a domestic cloud sidesteps per-token API bills. On risk, procurement, legal, and security teams will hesitate over provenance, licensing terms, and data-governance optics, especially for regulated sectors and public-sector work where METI-adjacent sensitivities and supply-chain scrutiny run high.
The pragmatic path for Japanese SIers is to treat models as swappable components behind an abstraction layer, benchmarking Hy4-class open weights against domestic and Western options on Japanese-language tasks, then deploying the cheapest model that clears the accuracy and compliance bar per use case. RPA and internal dev teams stand to gain most: a long-context open model self-hosted on local infrastructure can automate contract review and legacy-code comprehension without exporting sensitive data. The differentiator for Japanese integrators will not be which model they pick, but how quickly they build the evaluation, guardrail, and switching infrastructure that lets clients ride a rapidly commoditizing model layer without lock-in.
