Tencent open-sources Hy4 preview at 770B params and 1M context

Tencent's August 28 post announced Hy4 preview, an open-source language model with 770 billion total parameters and a context window of more than a million tokens, about four times the 256K that Tencent gives for Hy3. The total describes the whole model; Tencent's Hy4 preview model page says 49 billion parameters are active for each token, because each token is routed through 8 of 256 expert sub-networks plus one shared one. That keeps the computation per token far below the full size, but self-hosting still means storing every weight, which Simon Willison put at 1.56TB on Hugging Face (his note). The model page lists the license as Apache 2.0, the same permissive terms as Hy3.

The post's other claim, that Hy4 preview ranks "among the top tier of open-source models," comes with no named ranking in the post's text. The comparison Tencent does publish, on the model page and in its release announcement, is its own: 163 internal experts rated outputs on 203 engineering tasks without knowing which model wrote them, and Hy4 preview averaged 2.99 out of 4, against 2.92 for Zhipu's GLM-5.3 and 2.94 for Kimi K3. That is a narrow lead in a test Tencent designed and ran, and the model page also lists known problems, including reasoning longer than needed and over-checking its own work. The post points developers to Tencent Cloud TokenHub, a Tencent Cloud service that sells pay-as-you-go or reserved access to Tencent's Hy models alongside third-party models such as GLM and DeepSeek. The size and context window are concrete; the ranking awaits outside testing.