You know what’s annoying about ChatGPT? It forgets. Feed it a long document and ten minutes later it’s re-reading the same paragraphs. Tencent Hy4 is a new open-source model that holds a million tokens in memory at once, and it’s free to try right now.
What is Tencent Hy4?
Tencent Hy4 (the full name is Hy4 preview) is the latest large language model from Tencent’s Hunyuan team, released and open-sourced on August 28. It’s huge: 770 billion total parameters, with 49 billion active per request. That “active” number matters, because it means the model only switches on a fraction of its brain for any single task. You get the power of a giant model at a fraction of the computing cost. That’s the same trick used by other top open-weight models, and it’s why a model this large can still run affordably.
The context window is where things get wild. Hy4 handles over 1 million tokens. For comparison, its predecessor Hy3 handled 256,000, and that already felt generous. A million tokens is roughly the size of three huge novels, an entire codebase, or a year of meeting notes. You can drop all of it in one prompt.
Tencent says Hy4 ranks among the top tier of open-source models, and they put a number on it. In an internal blind evaluation with 163 experts and 203 engineering tasks, Hy4 scored 2.99 out of 4.00, slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. Take internal benchmarks with a grain of salt, sure. But the direction of travel is clear.
Why a 1M-token context matters
Context window is the model’s working memory. When you chat with most AI, it can only “see” the last chunk of conversation. Anything older gets summarized or dropped, and that’s when the model starts forgetting details you gave it an hour ago.
A 1M-token window changes what you can do in one session. You can paste a whole book and ask questions about chapter 3 with full knowledge of chapters 1 through 12. Feed it an entire repo and it can trace a bug across thirty files. You can even upload a year of support tickets and ask for patterns nobody had time to find.
For a beginner, the practical version is simpler: you stop having to chop your work into bite-size chunks. The model remembers, so your prompts can get lazier in the good way. Give it everything, once, and it keeps it straight.
What Hy4 is actually good at
Tencent built Hy4 for real-world productivity, not just benchmarks. That’s visible in the training data, which was co-created with Tencent’s own experts in software engineering, gaming, finance, and security. Here’s where it shows.
Coding
Hy4 is strongest at long-context development. That means understanding a big codebase, planning changes, debugging, and validating fixes across many files. Tencent also claims a specific trick: it can generate a playable game prototype from a single natural-language request, then keep refining it through conversation. If you’ve ever wanted to describe a game in plain English and watch it build, that’s the party trick.
Office work and research
The office side is where beginners will feel it most. Hy4 handles document creation, spreadsheets, and presentations, and it’s specifically tuned for data analysis and cross-document work. In scientific research, Tencent points to gains in molecular dynamics simulation, condensed-matter physics, and math. Wildly above most readers’ needs, but it explains why researchers are paying attention.
One more detail that’s genuinely interesting: Hy4 reportedly participated in its own development, proposing experiments for its own training and iterating on the results. A model that helps improve the next model. That’s either inspiring or mildly unsettling, depending on your mood.
Why the two thinking modes matter
Simon Willison’s teardown of the model found something worth knowing for practical use: Hy4 only has two reasoning settings, “high” and “no_think.” There’s no dial with eleven positions, which is honestly refreshing.
High is the default, and it’s what you want for anything that needs real thought: debugging code, analyzing a contract, planning a multi-step task. The model works through the problem before answering, and the quality shows on hard questions.
No_think means the model answers without the extended internal reasoning. That sounds like a downgrade until you remember the 1M-token context window. Big-context tasks are often retrieval tasks: “find every clause about termination in this contract” or “what did the Q3 report say about inventory?” For those, you don’t need the model to reason, you need it to search quickly. No_think will get you there faster, and since hidden reasoning tokens aren’t generated, it runs cheaper too.
So the practical tip for beginners: default to high for writing and problem-solving, switch to no_think when you’re mostly hunting for information inside a big document.
How Hy4 compares to other free AI models
The closest points of comparison for a beginner are GLM-5.3 and Qwen 3.8, both of which we’ve covered before. Here’s the side-by-side using Tencent’s own blind evaluation:
| Model | Company | Tencent blind-test score (/4.00) | Free to try? |
|---|---|---|---|
| Tencent Hy4 preview | Tencent | 2.99 | Yes: free for 2 weeks on WorkBuddy and CodeBuddy; OpenRouter from about $0.83 per million input tokens |
| GLM-5.3 | Zhipu AI | 2.92 | Yes: free tier on z.ai |
| Kimi K3 | Moonshot AI | 2.94 | Yes: via the Kimi app |
The honest reading: they’re all in the same band. Hy4 edges ahead in Tencent’s test, and it brings the biggest context window of the three. But it’s not a knockout. Choose based on what you need to do, not just the scoreboard.
How to try Tencent Hy4 today
You don’t need a server or a degree to touch this thing. Three routes, easiest first.
- Tencent’s apps. WorkBuddy and CodeBuddy are Tencent’s AI work apps, and Hy4 preview is free on both for two weeks after launch. That’s the zero-friction path. Yuanbao and ima also carry it.
- OpenRouter. If you don’t want another app, Hy4 is on OpenRouter. It costs about $0.83 per million input tokens and $2.50 per million output tokens, which works out to pennies per real session. Our OpenRouter tutorial explains the whole setup in about ten minutes.
- Run it yourself. The weights are open on Hugging Face, which means anyone with the hardware can host it. At 1.56 terabytes, that’s not a laptop project. Skip it unless you already run local models.
The honest catch
Three things to keep in mind.
First, it’s a preview. The “preview” in Hy4 preview is doing real work. Expect rough edges, and expect the final version to change. Second, the free window is a launch promotion, not a promise. Two weeks on WorkBuddy and CodeBuddy, then the apps move to their regular pricing. The open weights stay open, but “free” means “free to try right now.” Third, Tencent’s benchmark is Tencent’s benchmark. Independent testing will take weeks, and the leaderboards always shuffle.
Even with those caveats, the pattern here is worth noticing. The timeline from Hy3 in July to Hy4 in August is a one-month jump in context length and capability. Open-source models keep pulling ahead on memory, and the price of entry stays at zero. That’s the part that should matter to people who just want AI that actually remembers their documents.
If you want the full context on the free-model landscape, our GLM-5.3 explainer and the Qwen 3.8 review cover the other two serious players in this bracket.
