# Hy4 Defaults to Deep Reasoning. Tencent Says It Overthinks. **Summary:** Tencent's 770B open-weight model defaults to high reasoning even as its own model card warns of unnecessary reasoning and over-verification. I would test the mode before adopting the model. - Canonical: https://markhuang.ai/news/hy4-defaults-high-reasoning-overthinks - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-08-29 - Section: News - Tags: Tencent Hy4, Open Weights, AI Reasoning, Model Evaluation, AI Agents - Source: [Tencent](https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![A vast network of glass-like AI modules sends a glowing reasoning path through a repeated loop before reaching an exit](https://cdn.markhuang.ai/news/hy4-defaults-high-reasoning-overthinks/hero.webp) *Hy4 preview can route each token through a small share of a huge model. The harder product choice is deciding how much reasoning each task deserves.* [Tencent released Hy4 preview on August 28, 2026](https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/) with 770 billion total parameters, 49 billion activated per token, and a context window exceeding one million tokens. The weights are open, and Tencent is putting the model into products including WorkBuddy and CodeBuddy. The setting I care about is easy to miss in the launch numbers. Hy4 defaults to high reasoning, while Tencent's own model card warns that this preview can spend longer than necessary on complex tasks and over-verify its work. That tension can turn into latency, extra token use, and an agent that keeps working after it should have stopped. Teams need to test the reasoning mode as carefully as they test the model. ## The default carries a product opinion The [official Hy4 preview README](https://github.com/Tencent-Hunyuan/Hy4-preview/blob/main/README.md) describes high reasoning as the default and recommends it for math, coding, and reasoning. The same instructions expose a `no_think` option for direct responses. Tencent has therefore shipped two meaningfully different operating modes, then chosen the more deliberative one as the starting point. That choice may flatter a model during difficult evaluations. It is less obviously right for every production request. A coding agent sometimes needs to plan across a repository. It also needs to answer a narrow question, repair a format, or stop once the tests pass. More thought is useful only while it changes the result. I would treat reasoning effort as part of the product configuration, not as a personality trait of the model. A default decides what users pay and how long they wait before anyone has measured whether the extra work helped. ## Tencent's evidence earns a trial Tencent's blind internal evaluation asked 163 experts to rate outputs on 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Tencent also reports a 46.8% win rate against GLM-5.3 and 51.2% against Kimi K3, with ties and losses reported separately. Those results are specific enough to be useful, but they remain Tencent's evaluation with Tencent's experts and task mix. The narrow margins matter too. This is evidence that Hy4 belongs in a serious trial, not evidence that it should replace another model without workload testing. [BenchmarkList's third-party index](https://benchmarklist.com/models/tencent-hy4-preview/) currently tracks 40 Hy4 preview results. It records some of Hy4's strongest placements on codebase and long-context tasks, along with lower placements on other agent and software engineering tests. In an early [LocalLLaMA discussion](https://www.reddit.com/r/LocalLLaMA/comments/1w0igxk/tencenthy4preview_770ba49b_weight_dropped/), one commenter made the same practical objection, pointing to tests such as DeepSWE and CyberGym where the headline did not settle the comparison. > **Info:** > > I would run the same routine and difficult tasks in high reasoning and `no_think` modes, then compare accepted results, elapsed time, total tokens, tool calls, and human corrections. The winning mode is the cheapest one that clears the task reliably. ## Open weights make the default negotiable The release package is substantial. Tencent publishes the model under Apache 2.0, provides full and FP8 weights, and documents serving through vLLM and SGLang. The [Hugging Face repository](https://huggingface.co/tencent/Hy4-preview/tree/main) lists the full checkpoint at 1.56 TB across 131 weight shards. This is open access, but it is not a casual laptop download. Teams with the infrastructure can inspect the model card, choose the serving stack, tune the reasoning mode, fine-tune the model, and build their own evaluation harness. Everyone else will probably meet Hy4 through a hosted product or API. In that case, the provider's defaults and pricing shape the experience. This carries forward the question from my earlier look at [Hy3 making price part of the evaluation](https://markhuang.ai/news/hy3-price-is-the-eval). Hy4 is much larger, but the decision is still operational. A capable response that arrives after needless verification can be worse than a slightly weaker answer that finishes cleanly. As I argued in [Token Price Is Not the Bill](https://markhuang.ai/news/token-price-is-not-the-bill), cost belongs to the completed task, not the rate card or parameter count. ## A preview should have to earn its default I like that Tencent names the problem plainly. The model card calls Hy4 an early version and says the team expects more pre-training and post-training work. It also tells users what currently breaks: reasoning can run long, and verification can become excessive. Evaluators now have a concrete failure mode to watch. Disclosure does not fix the default. If high reasoning stays on for everything, the model's known weakness becomes the first experience most users receive. I would route high reasoning only to tasks that benefit from it. Hy4 preview looks strong enough to test and open enough to adapt. I would start by asking a less glamorous question than whether 770 billion parameters win the model race: on which tasks does Hy4 know when it is done?