What changed
On August 28, Tencent released and open-sourced Hy4 preview, a mixture-of-experts model with 770B total parameters, 49B activated parameters, and a context window reaching 1M tokens. The launch is available through Tencent’s announcement and the official Hy4 repository.
The notable part is the distribution plan around the weights. Hy4 preview is available inside Tencent’s WorkBuddy and CodeBuddy products, with a two-week free window at launch, and through Tencent Cloud TokenHub as an API. The official announcement lists pricing of $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens.
Why the product shape matters
A model release normally leaves builders to solve the adoption path themselves. Hy4 packages three paths together:
- Try it in a finished product. WorkBuddy and CodeBuddy give people a quick way to test the model against real office and coding workflows.
- Integrate it as a service. TokenHub exposes an OpenAI-compatible interface, so an existing client can point at Hy4 without adopting a new application protocol.
- Own the runtime. The repository includes the model weights, an Apache 2.0 license, finetuning and quantization guidance, and serving recipes for vLLM and SGLang.
That sequence turns model evaluation into a product funnel: experience the behavior, measure it in an application, then decide whether self-hosting is worth the operational cost. This is an inference from the launch packaging, but it is the useful distinction for builders: the model is being shipped with a route to adoption, not only a benchmark table.
The practical trade-off
The 1M-token context is useful for long documents, codebases, and multi-file analysis, but it does not make self-hosting lightweight. The official vLLM recipe uses the FP8 checkpoint with eight-way tensor parallelism, which puts Hy4 in serious infrastructure territory for most small teams. TokenHub’s model list makes the hosted contract more concrete: 1M context, up to 960K input tokens, up to 64K output tokens, structured output, function calling, reasoning, and cache support. Its calling guide also documents an OpenAI-style endpoint and tool-calling request.
There is another reason to keep the preview label visible. Tencent’s own repository says Hy4 can spend longer than necessary reasoning through complex tasks and can over-verify its work. A long context window is therefore a capacity, not a guarantee of lower latency or better product behavior.
Builder takeaway
For a product team, the sensible first experiment is hosted: route a bounded workflow through the API, measure answer quality, tool-call reliability, cache-hit rate, latency, and total cost, then compare those results with the hardware and maintenance cost of self-hosting. Keep the provider behind an adapter, and do not send an entire repository merely because the context window allows it.
The broader lesson is that open-model competition is becoming a packaging problem as much as a weights problem. The models that travel furthest will offer a credible path from first try to API integration to controlled deployment.