Tencent's new Hunyuan Hy4 takes the top spot on CWE-Bench-AA at 76% and sits on the capability-vs-cost Pareto frontier despite being trained with offensive security data stripped out.
- Tencent Hunyuan Hy4 ranks #3 on Artificial Analysis Cyber Index (58%) and #1 on CWE-Bench-AA (76%).
- Beats Grok 4.7 by 2 points at roughly 1/11th the cost, sits on capability-vs-cost Pareto frontier.
- API pricing is $0.834 input and $2.501 output per million tokens, averaging $0.92 per benchmark task.
- Uses 64k output tokens per task, 43% fewer than DeepSeek V4.1 Flash 1017.
- 770B-parameter MoE with vision and computer use, available on Tencent Cloud and OpenRouter.
- Trained with offensive cyber data (exploits, reverse engineering, pentest) deliberately filtered out.
Tencent’s Hy4 scored 58% on the Artificial Analysis Cyber Index, placing third overall and first on its vulnerability-patching sub-benchmark. Tencent achieved the result after excluding offensive security material from training, while charging an average of $0.92 per benchmark task.
Hy4’s edge is vulnerability patching #
The Cyber Index measures whether AI agents can find vulnerabilities in real repositories, reproduce the problems, and produce patches without breaking existing functionality. It gives equal weight to CWE-Bench-AA, DeepsecBench-AA, and CyberGym-E2E-AA.
| Hy4’s benchmark results | |||
|---|---|---|---|
| Evaluation | Score | Rank | Comparison |
| --- | --- | --- | --- |
| Cyber Index | 58% | 3rd | Behind GPT-6 Sol at 69% and Kimi K3.1 Pro at 64%; ahead of DeepSeek V4.1 Flash 1017 at 57% |
| CWE-Bench-AA | 76% | 1st | Six points ahead of DeepSeek V4.1 Flash |
CWE-Bench-AA is Artificial Analysis’s implementation of Collinear AI’s CWE-bench. It evaluates whether an agent can identify and repair real software vulnerabilities, making the result more relevant to repository-level remediation than tests based on describing vulnerability categories.
High scores at $0.92 per task #
Hy4 costs $0.834 per million input tokens and $2.501 per million output tokens. Its average Cyber Index task costs $0.92, placing the model on the benchmark’s capability-cost Pareto frontier. No evaluated model delivers both a higher score and a lower task cost.
| Hy4 against nearby competitors | ||
|---|---|---|
| Model | Score difference | Cost and token comparison |
| --- | --- | --- |
| Grok 4.7 (xhigh) | Hy4 leads by 2 points | Hy4 costs roughly one-eleventh as much and uses 37% fewer output tokens |
| DeepSeek V4.1 Flash 1017 (max) | Hy4 leads by 1 point | Hy4 costs twice as much and uses 43% fewer output tokens |
| Kimi K3.1 Pro | Hy4 trails by 6 points | Hy4 uses 49% more output tokens |
| GPT-6 Sol (Daybreak Blue) | Hy4 trails by 11 points | Hy4 uses 56% more output tokens; GPT-6 Sol has trusted-access availability |
Hy4 averages about 64,000 output tokens per task. GPT-6 Luna remains cheaper at $0.12 per task, while Hy4 offers a higher-scoring option for teams balancing defensive capability against API spending.
Defense-only training holds up #
Tencent’s reported training policy included defensive material related to vulnerability remediation and proof-of-concept generation, meaning code that demonstrates a flaw. It excluded exploit development, reverse engineering, and penetration-testing content.
Tencent’s internal security teams reported no elevated offensive-task risk during red-team testing. The public benchmark cannot independently validate that safety claim because its scope is defensive: agents inspect source code and repair vulnerabilities, while weaponizing flaws falls outside the evaluation.
The 76% CWE-Bench-AA score provides narrower evidence that removing offensive training content did not prevent strong performance on defensive repository work. It does not establish how Hy4 performs on exploit development or other offensive tasks.
A 770-billion-parameter model #
Hy4 is Tencent Hunyuan’s flagship 770-billion-parameter mixture-of-experts model. This architecture routes each token through a subset of specialized parameters instead of activating the entire model. The full release builds on the preview issued in August and adds vision and computer-use capabilities.
Developers can access Hy4 through Tencent Cloud and OpenRouter at the preview model’s API price. Tencent has announced downloadable weights and system cards for the full release, though neither was public at the time of publication. The preview weights are available on Hugging Face under the Apache 2.0 license.
Where Hy4 fits in a security stack #
Hy4’s benchmark profile makes it a candidate for defensive workflows that operate directly on source repositories, including:
- Triaging reported vulnerabilities
- Generating and validating patches
- Following up on static application security testing findings
- Investigating security-relevant code paths
Teams whose workflows include exploit development, reverse engineering, or red-team tooling should evaluate those tasks separately. Hy4’s filtered training data and the defensive scope of the Cyber Index leave its offensive capabilities largely unmeasured.
The Cyber Index also remains separate from the general Artificial Analysis Intelligence Index. Hy4’s result therefore provides limited evidence about knowledge work, long-context processing, coding outside security, or multimodal computer use. Those workloads require their own evaluations.
Hy4 offers a specific trade-off: competitive repository-level remediation at a low API cost, paired with limited public evidence about offensive security and broader model capabilities. The pending weights and system cards will determine how thoroughly developers can inspect, deploy, and test the full release.