Hy4 Uncensored: Specs, Pricing and Best Uses

What's new in Hy4?

Images and longer work. Tencent says Hy4 is its first model with visual understanding, covering image processing, chart analysis and visual math. It also adds computer use, for both command-line and graphical apps, along with better instruction following, long-context reading and long agent tasks. In Tencent's internal tests, the hallucination rate fell from 12.95% on the preview to 7.2%.

On Silk Compute, Hy4 reads up to 1M tokens of context and writes up to 66K tokens in one reply, with reasoning effort you can set to low, high or max.

How does Hy4 compare with our other models?

On Vals AI's tests it runs close to GLM-5.3, at a lower price, and it tops Tencent's own finance panel. Kimi K3.1 Pro stays far ahead on terminal work.

The first four rows are Vals AI's own runs, as percentages from its model pages, read November 5, 2026, with margins of about one to three points, so treat small gaps as ties. The finance panel is Tencent's own blind test, with 213 financial professionals on 579 tasks. Kimi is K3.1 Pro and DeepSeek is V4.1 Flash 1017. Every score is for the original model, not the abliterated one.

Per 1M tokens, Hy4 costs $1.60 / $0.30 / $3.20 for input, cached input and output, against $2.00 / $0.30 / $4.00 for GLM-5.3.

Is Hy4 good for security work?

For defensive work, yes. For authorized exploit development beyond a proof of concept, use Kimi K3.1 Pro. Tencent says it trained Hy4 on defensive security data, with a focus on proof-of-concept development and fixing vulnerabilities, and that "offensive materials related to exploitation are explicitly excluded from the training dataset." Its security labs also "found no significant concerns regarding Hy4's offensive cybersecurity capabilities in the areas evaluated." Removing refusals can't add training data, so we'd expect Hy4 to trail on exploit work. That's our reading, not a benchmark.

So give Hy4 the screenshots, dashboards and diagrams from an investigation, and the fixes that follow. Kimi K3.1 Pro is the best model for security research and red teaming, and our security research guide has the setup.

Which jobs should you give it?

Our coding agent guide goes deeper on choosing a model for an agent.

What should you know before calling it?

Send images as base64 next to your text, in any of our three API formats. The models docs have the field names. Images count as input tokens. Check anything Hy4 reads off an image: in our small test on images of numbers, it got 8 of 9 right.

Leave plenty of room in max_tokens, because Hy4 thinks at length and thinking counts toward the limit. Tencent's preview card lists "spending longer than necessary reasoning through complex tasks" as a known issue. In our runs, a 6,000-token limit on a word problem went entirely on thinking at low, high and max, and the reply came back with no text. Set the limit well above the answer you expect. The reasoning docs have more.

Can you run Hy4 yourself?

Not the final version yet. Tencent says it plans to publish the weights and system cards "in the coming days." The preview is out under Apache 2.0 with 770B parameters, and its 8-bit file alone is 814 GB, so plan on a rack. Our abliterated AI size chart shows what does fit.

Specs and pricing

How do I get started?

Create an API key and send a request to hy4-uncensored at https://api.silkcompute.com. The model page has a prompt you can paste into your coding agent to set it up for you, plus code for Python, JavaScript, cURL, Claude Code, Codex and OpenCode.