Hy4

General release of Tencent Hy Team's fourth-generation flagship, shipped on November 4, ten weeks after the open-weight Hy4 preview and six days after Dowson Tong said at Tencent's Global Digital Ecosystem Conference that it was coming. It is still a 770B-parameter MoE, the same total as the preview. Tencent says pre-training and post-training were both expanded, without giving token counts. The line gains two new capabilities: visual understanding (fine-grained image processing, chart analysis, math problems posed as images) and computer use that mixes command-line and GUI operation. Tencent says it trained specifically on Blender, CAD and long cross-application tasks, and validated in a sandbox of hundreds of applications. Tencent claims the model has entered the top tier of open-source models, and Bloomberg's headline reports the claim as ranking among China's best. Tencent itself acknowledges "a significant gap" to the top proprietary models.

All results are Tencent's own, mostly internal evaluations. Against the preview: GameDev +19.1%, GameLogic +9.8%, financial hallucination rate 12.95% → 7.2%, guessing-error rate 23.1% → 5.6%. In a financial crowdsourced test (213 experts, 579 tasks), raters gave Hy4 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM 5.3. A WorkBuddy office evaluation gave an average of 83.18 and a 96.15% task pass rate. Tencent reports 30.4/100 on its own Hy-BioExperimentBench and calls this a significant gap to frontier models. Tencent's Xuanwu and Keen labs audited cyber capability and found no significant offensive uplift, after offensive security content was excluded from training. Tencent also says the model found training-system bottlenecks that raised its own training throughput by 20.1%.

The model is live in Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, Yuanbao, ima and the WeCom assistant, and through the OpenRouter, Hermes and OpenCode APIs. List prices are unchanged from the preview at ¥6/¥18 per M input/output tokens (¥0.3 for cache hits), with TokenHub at 50% off through December 31. The weights are not yet released. Tencent says they "will be open-sourced soon" but gives no date; its English announcement says it "plans to release model weights and system cards in the coming days". As of 2026-11-10 the tencent HuggingFace org holds no Hy4 repository beyond the preview. A cheaper Hy4 flash is "scheduled for release soon" and had not shipped as of the same date. Artificial Analysis has since scored it 35 on the Intelligence Index (AA v5.0.1), just above Inkling 2 Small's 31.