Tencent Open-Sources Hy4, a 770-Billion-Parameter Model Built for Office Work
Tencent released Hy4 preview with 770B total parameters, 49B active, and a context window above one million tokens. The unusual part is what it was optimised for: spreadsheets, documents and long coding sessions.
Tencent has released and open-sourced Hy4 preview, the next generation of its Hunyuan model family. The numbers are large: 770 billion total parameters, 49 billion of them active per request, and a context window above one million tokens. Parameters are the adjustable values a model learns during training, roughly a measure of raw capacity. The gap between total and active is the mixture of experts trick, where the model holds many specialised sub-networks but only wakes a fraction of them for any given request, which keeps the running cost far below what the headline size suggests. Context window means how much text the model can hold in mind at once, and a million tokens is roughly a small shelf of books.
What stands out is the target. Tencent did not pitch Hy4 at benchmark leaderboards but at what it calls real-world productivity: coding across long sessions, spreadsheets and documents, financial analysis, and scientific research. In a blind internal evaluation with 163 experts across 203 engineering tasks, Hy4 preview averaged 2.99 out of 4.00, slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. Tencent also says the model took part in its own development, proposing and running experiments on training methods and low-level code, and that it analysed its own inference bottlenecks well enough to lift end-to-end throughput by 31.8 percent. API pricing is 0.834 dollars per million input tokens and 2.501 dollars per million output tokens, with cache hits at 0.042 dollars.
What’s actually going on here: open weights means the trained model files are published, so anyone with enough hardware can download and run it themselves instead of calling an API. Chinese labs have made this their competitive strategy, and the pace has become relentless. Hy4 lands the same week as GLM-5.3’s weights, and days after Qwen and IBM releases. Every one of these drops pushes down the price of capable AI for everyone, including customers of the closed labs, because those labs have to price against something that is free to self-host. Worth keeping expectations grounded on the self-improvement claim: a model suggesting optimisations to its own inference code is real and useful engineering, but it is a long way from a system that redesigns itself.
What this means for you: almost nobody runs a 770B model at home, so for most people the practical effect arrives indirectly, through cheaper API prices and through smaller distilled versions that follow later. If you are curious, Hy4 is free to try for two weeks through Tencent’s WorkBuddy and CodeBuddy, and reachable via OpenRouter. If you already pay for an AI coding subscription, the useful move is not to switch today but to keep an eye on price per token over the next quarter, because this is exactly the kind of release that moves it. And if data residency matters to you, note that these are Chinese cloud products, while the open weights themselves can be run wherever you like.
Sources
Source: https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
Beatport Bans Fully AI-Made Tracks, and Publishes the Numbers Behind the Decision
The DJ marketplace now blocks tracks made entirely or mostly by AI, screening uploads with a detection tool from Beatdapp. A Beatport survey found 60 percent of users would not play AI music in a set.
