On Friday, Tencent posted a preview of a new open model to Hugging Face, aimed squarely at software engineering, research and financial analysis. Reuters, which reported the release first, described it as a mixture-of-experts system with roughly 770 billion total parameters, of which only about 49 billion fire on any given request. Tencent published a smaller, quantized version alongside it and seeded the files across several platforms at once. Anyone can download the weights and run them for free. The interesting question is why a company would give away something that expensive to build.
The answer is in the detail Reuters flagged almost in passing: Tencent plans to wire the model into products it already sells, developer and productivity tools it markets under names like CodeBuddy and WorkBuddy. That reframes the release. The open weights are not a gift to the community so much as the engine inside paid software. Give away the model, charge for the tools that make it useful and the cloud that runs it at scale. It is the same logic a growing list of Chinese labs have settled on, and it explains why "open" has become a competitive weapon rather than a philosophy.
The mixture-of-experts design is central to making that math work. Instead of running every one of its 770 billion parameters for each query, the model routes each request through a small set of specialist sub-networks. You get the breadth of a very large model at the compute cost of a much smaller one. This is not a Tencent invention. DeepSeek, Alibaba's Qwen team and Zhipu's GLM line have all converged on the same architecture for their biggest systems, because it is the only way to keep inference cheap enough to hand out for free. The convergence is itself the story: a whole national industry has agreed on how to build these things and on how to make money from them.
Tencent's own preview materials make bolder claims, and these deserve more caution. In internal blind evaluations it says it ran, involving 163 experts scoring 203 engineering tasks, the model reportedly averaged 2.99 out of 4, edging past Zhipu's GLM-5.3 and Moonshot's Kimi K3. The company also touts a context window beyond a million tokens and a party trick of generating a playable game prototype from a single sentence. Self-reported benchmarks from the vendor that built the model are worth exactly what independent testing later confirms, which so far is nothing. The naming has been muddy too, with coverage variously calling this the Hy3 and Hy4 preview.
Strip away the marketing and the shape of the thing is clear enough. A major Chinese platform has put a frontier-scale coding model into the open, tied it to software it already sells, and priced the model itself at zero. For developers outside China, it is another capable option that costs nothing to try. For the American labs preparing to sell premium access to closed models, it is one more data point in a trend that keeps pushing the default price of raw capability toward the floor. The weights are free. The money, as ever, is moving to whoever owns the workflow.