Image: METAL
Summary
- Tencent Hunyuan 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 it 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 says the model took part directly in optimizing its own training methods and inference infrastructure, raising 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. Tencent says it scored 65.7% on SWE-bench Pro, a coding and software-engineering benchmark, edging out OpenAI's GPT-5.6 Sol (64.6%) and Zhipu AI's GLM-5.3 (64.6%) by a narrow margin.
The announcement lands in the middle of an intense race among Chinese open-source LLMs. On August 18, Zhipu AI released the 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 have also been released as open-source models around the same time.
Performance comparison
Alongside SWE-bench Pro, Tencent released results from its own blind evaluation. In that test, 163 experts graded the model on 203 real-world engineering tasks, and Hy4 preview scored 2.99 out of a possible 4.00, narrowly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). Worth keeping in mind: both sets of numbers come from Tencent's own 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 described the model as having "joined the ranks of top-tier open-source models."
The model reportedly helped build itself
The Hunyuan team says Hy4 preview took part directly in its own development process. According to Tencent, this marked the first time the model participated in automated optimization of training methods, data strategy, evaluation frameworks, and low-level compute operators — 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 on its own. That points to an early stage of a broader industry trend: AI models running recursive self-improvement experiments without direct help from human researchers.
Tencent says the model also analyzed bottlenecks in its own inference system and ran multiple rounds of operator fusion and communication optimization, lifting end-to-end throughput by 31.8% over baseline. That gain held consistently, Tencent added, across different context lengths and levels of concurrent requests.
How to try it
Hy4 preview is available now through the Cline coding agent, and it's also accessible through Tencent's WorkBuddy and CodeBuddy apps, as well as the Yuanbao and ima apps. For API access, users can go through Tencent Cloud TokenHub or OpenRouter.
- Select Hy4 preview from the model list in Cline or the CodeBuddy app.
- WorkBuddy and CodeBuddy offer it free for two weeks starting August 29, and free access to the previous model, Hy3, has been extended through September 30.
- For direct API integration, 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 can be downloaded and run on anyone's own servers worldwide, and the cloud app access is open globally as well. Tencent points to use cases including debugging and code review that require long context windows, office work spanning documents, spreadsheets, and presentations, and building playable game prototypes from a single natural-language request.
According to domestic reporting, Tencent Cloud has been courting Korean businesses with practical, low-cost AI rather than top-tier flagship models, and Hy4 preview's low API pricing and open-source release fit that same strategy.
Editor's view
Chinese open-source developers have moved past simply claiming to match GPT-tier models on a few benchmark points — they're now letting models tinker with their own training and inference infrastructure. Hy4 preview's self-improvement claims come with the caveat that they're still early-stage, but the fact that handing training-pipeline optimization to the model itself, rather than to human engineers, is showing up in an actual product announcement is itself a signal that this trend has moved out of the lab.
There's an interesting generational contrast here too. Extending free access to the previous model, Hy3, through September 30 while also opening the new Hy4 for two weeks free reflects a preview-first strategy: real user feedback gets folded straight into the next training round. This isn't the first time Hunyuan has released a preview first and polished it into a final version later — it's a pattern the team has repeated.
From a practical standpoint, there are two things to weigh separately. One is the benchmark numbers — both SWE-bench Pro and the internal blind evaluation come from Tencent's own measurements, so they should be treated as pending independent verification. The other is price: at $0.834 per million input tokens, the cost could be low enough to matter for teams running coding agents constantly, which makes it worth testing directly through Cline or CodeBuddy in a real workflow.
Tencent has already signaled that "the next batch in the Hy4 series is coming soon," so a full release or another preview could follow within weeks. With the benchmark gap against GLM-5.3 and Kimi K3 now down to one or two percentage points, the next round is likely to be less about chasing numbers and more about how mature that self-improvement loop actually turns out to be.
Comments