Tencent's Hy4 adds vision and computer use; full weights promised in days, preview already open

Tencent announced Hy4 on November 4: the latest Hunyuan-line flagship, now with visual understanding and computer use for the first time in that family. The model is live on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy and several other Tencent products, and through third-party APIs such as OpenRouter. Pricing stays where Hy4 Preview sat — about $0.90 / $2.70 per million input/output tokens, $0.045 for cached input — with a 50% TokenHub discount through December 31.

One careful distinction: Tencent calls Hy4 one of the world's leading open-source models, but the November 4 note says full model weights and system cards will come "in the coming days." What is already on Hugging Face under Apache 2.0 is Hy4 preview (and an FP8 build): a 770B-total / 49B-active MoE with a 1M-token context, Gated DeepSeek Sparse Attention, identity Hyper-Connections, and a native MTP draft layer. The preview card itself still lists known limits — over-long reasoning and a tendency to over-verify its own work.

What Tencent claims changed

Beyond vision (fine-grained image work, charts, visual math) and computer use (CLI and GUI, with extra training on Blender and CAD), the company reports product-tied gains: WorkBuddy averages 83.18 with a 96.15% pass rate across twelve mainland-China scenarios; GameDev and GameLogic internal scores rise 19.1 and 9.8 points over Hy4 Preview; an internal hallucination rate falls from 12.95% to 7.2%. A blind financial eval with 213 professionals across 579 tasks puts Hy4 at 3.04 against DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 — company-run numbers, not an independent leaderboard.

On safety, Tencent says Hy4 was trained on defensive security data (PoCs and remediation) with offensive exploitation material excluded, and that Xuanwu Lab and Keen Security Lab found no significant offensive-capability concerns in the areas they checked. Biosafety review, per the same post, finds strong foundational biology knowledge but weaker open-ended experimental design than frontier closed models.

Hy4 Flash is promised soon for cheaper high-volume use. The open question for a follow-up is whether the full Hy4 weights match the API model the announcement describes, and how they compare with Hy4 preview once independent benches land.