Hy4 model: Tencent ships the API before weights, parameter count or license

Tencent's announcement headline says Hy4 joins the ranks of the world's leading open-source AI models. That label cannot be checked until the weights and license terms are out. Until then, engineers can call the model through an API but cannot plan a self-hosted deployment, because the inputs a serving plan needs are missing: model size, memory footprint and license.

The announcement describes capabilities and gives no specifications. Tencent says Hy4 introduces visual understanding for the first time, citing fine-grained image processing, complex chart analysis and mathematical visual reasoning. It also introduces computer use for the first time, covering both command-line and graphical user interfaces. Tencent says the model received specialized training on professional software such as Blender and CAD to improve planning and execution of long-horizon tasks. The release claims better instruction following, long-context comprehension and long-horizon agentic tasks, but gives no context-window length.

Hy4 is available in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant (Dayuan) and ima. Its API is also being made available through OpenRouter, Hermes and OpenCode. Tencent says integration with QQ, Honor of Kings, Game of Peace, WeRead and Tencent News is underway.

The launch post gives no parameter count. It does not say whether the model is dense or mixture-of-experts, and gives no active-parameter figure, no context length and no training-compute figure. The only mention of architecture is indirect: Tencent says Hy4, using the Hyra 2.0 harness, optimized the pre-training infrastructure for smaller models sharing the same architecture. So smaller models with Hy4's architecture exist, but the architecture itself is not described.

The license is also unstated. 'Open-source' could mean a permissive license, a custom license with use restrictions, or weights released without training code or data, and the post does not say which. Teams that need license terms for a procurement or compliance decision have to wait for the weights release.

The launch material gives no GPU count, VRAM requirement, quantization guidance or throughput figure. Without a parameter count and architecture, nobody can estimate weight memory, KV-cache cost per token or the minimum node size. Any figure published before the weights are released is a guess.

The API price is the one cost input available. Tencent lists about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, the same as Hy4 Preview. Hy4 is 50% off standard pricing on Tencent Cloud TokenHub through December 31, 2026, and WorkBuddy and CodeBuddy are free from November 4 to 15.

Tencent says Hy4 Flash, which it describes as delivering greater cost-efficiency, is scheduled for release soon. The post's own section heading says Hy4 Flash was 'Recently Released', so its status is unclear. No price or specification has been given for it.

The Tencent Hy team and Tencent Financial Technology ran the comparison as a blind evaluation. It involved 213 financial professionals, including 153 experts from external institutions, and about 70 percent of them had more than five years of industry experience. Scores were based on performance across 579 real-world financial tasks:

Several caveats apply. Tencent does not state the scoring scale, so the 0.07 gap to DeepSeek cannot be judged as large or small. The DeepSeek entry is DeepSeek-V4.1-Flash-1017 and the Kimi entry is Kimi K3.1 Pro, so the results may not carry over to other variants of those models. The test covers a single domain and was organized by Tencent's own teams.

The announcement reports no standard public benchmark scores, such as coding, math or general reasoning suites, for any of these models. Mistral Large 4 is absent entirely, so the launch material offers no comparison with it.

The remaining numbers compare Hy4 with Hy4 Preview or report scores on Tencent's own products. They show progress inside the Hy4 line but do not position it against other models:

For agent builders, the drop in unsupported-assumption errors from 23.1% to 5.6% is likely the most relevant figure, because errors like this can propagate across multi-step tool use. Tencent reports it alongside targeted search optimization aimed at retrieving up-to-date information, integrating multiple sources and verifying factual consistency. It is still an internal metric with no published methodology. Tencent has also open-sourced two foundational models developed with Hy4 and Hyra 2.0: HyCell for single-cell representations and HyPDE for physics fields.

Tencent says Hy4 was trained on defensive security data, prioritizing proof-of-concept development and vulnerability remediation. The training data includes reverse engineering and penetration testing for analysis purposes, and offensive exploitation material was explicitly excluded. Experts from Tencent's Xuanwu Lab and Keen Security Lab assessed cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation, and found no significant concerns in the areas evaluated.

On biosafety, internal expert reviews found the model's foundational biological knowledge strong and its open-ended experimental design more limited than frontier models. Both assessments are internal. The system cards promised alongside the weights are where the methodology should appear, and teams with open-weight risk reviews should hold sign-off until those cards are published.