← Back to feed

Tencent released Hy4 on November 4, the general-availability follow-up to the August 28 preview. It keeps the 770B-parameter mixture-of-experts design (49B active, 1M-token context) and the preview’s API price, and adds two things the preview lacked: vision, and computer use across command-line and graphical interfaces.

Unlike the preview, which shipped with Apache 2.0 weights on day one, Hy4 launched API-only. Tencent says it will release the weights and a system card “in the coming days”, with a focus on cybersecurity and biological safety evaluations.

Key Numbers

• Architecture: 770B total, 49B active, 1M-token context
• Pricing per million tokens: $0.834 input, $2.501 output, $0.042 cached input (unchanged from preview)
• Discounts: 50% off on Tencent Cloud TokenHub through December 31; WorkBuddy and CodeBuddy free November 4 to 15
• Availability: Tencent Cloud TokenHub, OpenRouter, WorkBuddy, CodeBuddy, Yuanbao, ima
• Internal finance evaluation: 579 tasks, 213 professionals (153 external). Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68)
• Game development: +19.1% on GameDev and +9.8% on GameLogic versus the preview (baselines not published)

What changed from the preview

The August preview was text-only. Hy4 now interprets charts and fine-grained image details, and operates software on its own. Tencent says it trained specifically on professional tools such as Blender and CAD and on long-horizon workflows that span several applications. It also targets Unreal Engine, Unity and Godot, including multiplayer game logic, and says it can co-edit Word, Excel and PowerPoint files with a person in real time.

On the AI R&D side, Tencent says Hy4 paired with its Hyra 2.0 harness optimised internal pre-training infrastructure for smaller same-architecture models and trained a variant that beat an internal baseline on several benchmarks. It is open-sourcing two foundation models built that way: HyCell for single-cell representation and HyPDE for physical field modelling.

Reading the claims

Nearly everything above is Tencent’s own evaluation. The finance and research results (TenureTrail, 123 internal research tasks, where Hy4 ranks first among open-weight models) are internal benchmarks with Tencent’s choice of comparators. The one number that matters for the pricing story is the one not yet independently reproduced: how Hy4 sits against closed frontier models on standard agentic suites. Tencent acknowledged, as reported by the South China Morning Post, that the model still trails leading closed-source systems on research-level mathematics, long-horizon coding, self-improvement and advanced visual question answering.

For reference, the preview’s published numbers were Terminal-Bench 2.1 at 85.4% and SWE-bench Pro at 65.7%, against Claude Opus 5 at 79.2% on SWE-bench Pro. At $2.501 per million output tokens, Hy4 is priced as the cheap option in the frontier tier.

The open question

The weights are the story to watch. The preview’s Apache 2.0 licence was the reason Hy4 drew developer attention in August, and the same open playbook has been used all year by DeepSeek, Alibaba’s Qwen and Moonshot. A GA release that is API-only, even temporarily, means no one outside Tencent can verify the new vision and computer-use capabilities locally until the checkpoint ships. Tencent has also said a cheaper Hy4 Flash is coming soon.