On November 4, Tencent Hy published Hy4 on its official site. Versus Augusts Hy4 preview, Tencent says it scaled up pre-training and post-training, added vision for the first time, added computer use across CLI and GUI, and trained specifically on Blender, CAD and engines like Unreal, Unity and Godot. Hy4 is live on Tencent Cloud and OpenRouter (tencent/hy4, 1M context) at the same API price as the preview. Weights and a system card are promised soon, and a cheaper Hy4 flash is coming.
Key Takeaways
- ✓Official blog: Introducing Hy4, dated 2026-11-04; the Tencent Hy transparency center also lists V1.0 Hy4
- ✓Two firsts: vision (fine-grained image detail, complex charts, math visual reasoning) and computer use (CLI plus GUI, with targeted training on Blender, CAD and long cross-app workflows)
- ✓Tencent's own evals: on 579 real finance tasks rated by 213 finance professionals Hy4 scores 3.04, first among open-weight models (DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68, GLM-5.3 2.68); GameDev +19.1%, GameLogic +9.8%; internal search hallucination rate 15.76% to 5.45%
- ✓Same API price as the preview: $0.042 cached input, $0.834 input, $2.501 output per million tokens; the OpenRouter page currently shows 50% off ($0.417/$1.251), 1,048,576 context and up to 65,536 output tokens
- ✓Weights are not out yet: Tencent says it will open-source them with a full system card, working with vLLM and SGLang; Hy4 flash is due soon
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
August's Hy4 preview was text-only, and Tencent itself flagged known issues such as over-long reasoning and excessive self-verification, which limited it for agent tasks that need to see a screen. On November 4, Tencent Hy published Hy4 on its official site, calling it the first time Hy ranks among the strongest open-source models and pitching it at real-world productivity.
How it works
Per the official blog, Hy4 substantially scales both pre-training and post-training and adds vision for the first time, combining perception, reasoning and tool use for fine-grained image detail and complex charts. It also adds computer use across command-line and graphical interfaces, with targeted training on Blender, CAD and long cross-app workflows, plus adaptation for Unreal Engine, Unity and Godot in frontend and game work. The release blog does not restate parameter counts; Tencent's preview figures were 770B total, 49B active and a 1M context.
Benchmarks
All numbers below are Tencent or partner internal evals with no third-party reproduction yet. On 579 real finance tasks Hy4 scores 3.04, first among open-weight models (DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro and GLM-5.3 both 2.68). Versus the preview it gains 19.1% on GameDev and 9.8% on GameLogic, cuts the search hallucination rate from 15.76% to 5.45%, and scores 61.63 on RabyteTech's IRAB, 12.03 points higher. Tencent also admits a notable gap to top proprietary models on research-level math and long-horizon coding.
Getting started
Call it now on Tencent Cloud or OpenRouter (model ID tencent/hy4, with image input, tool calling and structured output), or try it in CodeBuddy, WorkBuddy, Yuanbao and ima. Official pricing matches the preview at $0.834 input and $2.501 output per million tokens; OpenRouter currently shows 50% off. Weights and a system card are coming "soon," so local deployment has to wait, and a cheaper Hy4 flash is on the way.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.