이미지: METAL LAB 생성
Summary
- Tencent's Hunyuan team has released an open-source LLM called Hy4 preview, with 770 billion total parameters, 49 billion active parameters, and a context window exceeding one million tokens.
- Tencent says the model scored 65.7% on SWE-bench Pro, edging out GPT-5.6 Sol (64.6%), GLM-5.3 (64.6%), Kimi K3 (63.3%), and DeepSeek V4 Pro (60.3%).
- Tencent also says the model took part directly in optimizing its own training methods and inference infrastructure, lifting throughput by 31.8%.
The model carries 770 billion total parameters, 49 billion active parameters, and a context window that stretches past one million tokens. On SWE-bench Pro, a benchmark for coding and software engineering tasks, Tencent says it scored 65.7%, narrowly ahead of OpenAI's GPT-5.6 Sol (64.6%) and Zhipu AI's GLM-5.3 (64.6%).
The announcement lands in the middle of an intense race among Chinese open-source LLMs. On August 18, Zhipu AI released its GLM-5.3 API, pushing its Terminal-Bench score from 4.6 to 28.3 through post-training alone. Moonshot AI's Kimi K3 and DeepSeek's V4 Pro are also out as open-source releases around the same time.
Benchmark comparisons
Alongside SWE-bench Pro, Tencent also released results from its own blind evaluation. In that test, 163 experts graded 203 real-world engineering tasks, and Hy4 preview scored 2.99 out of 4.00 — slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). It's worth keeping in mind that both sets of results come from Tencent's own internal measurements.
| Benchmark | Hy4 preview | GPT-5.6 Sol | GLM-5.3 | Kimi K3 | DeepSeek V4 Pro |
|---|---|---|---|---|---|
| SWE-bench Pro | 65.7% | 64.6% | 64.6% | 63.3% | 60.3% |
| Evaluation (out of 4.00) | Hy4 preview | GLM-5.3 | Kimi K3 |
|---|---|---|---|
| Expert blind evaluation | 2.99 | 2.92 | 2.94 |
Tencent's announcement describes the model as having "joined the ranks of top-tier open-source models."
The claim that the model helped build itself
The Hunyuan team says Hy4 preview took an active role in its own development process. For the first time, the model participated in optimizing training methods, data strategy, evaluation frameworks, and low-level operator automation — proposing and testing its own approaches, then feeding those results back into the next round of exploration. Tencent describes this as an early-stage recursive self-improvement loop.
In plainer terms, Tencent says Hy4 preview experimented with and improved its own training methods and inference speed. That points to something the industry is only just starting to explore: AI models improving their own performance without help from human researchers.
Tencent says the model also analyzed bottlenecks in its own inference system, running multiple rounds of operator fusion and communication optimization. The result, according to Tencent, was a 31.8% increase in end-to-end throughput over baseline, a gain that held consistent across different context lengths and concurrent request loads.
How to try it
You can try Hy4 preview right away in the Cline coding agent, or access it through Tencent's WorkBuddy and CodeBuddy apps, as well as the Yuanbao and ima apps. For API access, go through Tencent Cloud's TokenHub or OpenRouter.
- Select Hy4 preview from the model list in Cline or the CodeBuddy app.
- WorkBuddy and CodeBuddy are offering it free for two weeks starting August 29, and free access to the previous model, Hy3, has been extended through September 30.
- To connect via API directly, get a key through Tencent Cloud TokenHub or OpenRouter. Pricing runs $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens.
Tencent says the open-source weights are available for anyone worldwide to download and run on their own servers, and that the cloud app access is open globally as well. Suggested use cases from Tencent include long-context debugging and code review, office work that spans documents, spreadsheets, and presentations, and building playable game prototypes from a single natural-language request.
According to domestic reporting, Tencent Cloud has been targeting Korean businesses not with top-tier flagship models but with practical, low-cost AI — and the low API pricing and open-source release of Hy4 preview fit that same strategy.
Editor's take
Chinese open-source developers have moved past the stage of simply claiming they've matched GPT-class models by a few benchmark points. Now they're pushing into territory where the model itself helps tune its own training and inference infrastructure. Tencent's self-improvement loop claim comes with the caveat that it's still early-stage, but the fact that handing training-pipeline optimization to the model instead of human engineers has made it into an actual product announcement is itself a sign this idea has left the lab.
There's an interesting generational contrast here too. Extending free access to the previous model, Hy3, through September 30 while also offering the new Hy4 free for two weeks reflects a preview-first strategy: feed real user feedback straight into the next training round. Releasing a preview first and polishing toward a full launch later isn't new for Hunyuan — it's a pattern they've repeated before.
From a practical standpoint, there are two things worth separating. One is the benchmark numbers themselves — both SWE-bench Pro and the internal blind evaluation are Tencent's own measurements, so they should be treated as unverified until independently confirmed. The other is pricing: at $0.834 per million input tokens, the cost is low enough to meaningfully ease the burden for teams running coding agents constantly, which makes it worth plugging into an actual workflow through Cline or CodeBuddy to judge for yourself.
Tencent has already signaled that "the next batch in the Hy4 series is coming soon," so a full release or follow-up preview within weeks seems likely. With the benchmark gap against GLM-5.3 and Kimi K3 now down to roughly one to two percentage points, the next round will probably be less about squeezing out more benchmark points and more about how mature that self-improvement loop actually turns out to be.
Comments