Tencent Hy4: 1M-Context Model With Vision and Computer Use, Now on OpenRouter (Weights Pending)

Tencent Hy4 launched on November 4, 2026. It’s the full version of the Hy4 Preview model, and it adds two things the preview didn’t have: image understanding and computer use. You can call it today through OpenRouter as tencent/hy4, with a 1,048,576-token context window and up to 65,536 output tokens.

Tencent’s announcement says it plans to release Hy4 model weights and system cards “in the coming days”. As of November 9, 2026, there’s no Hy4 repository under Tencent’s Hugging Face organization. Only the earlier Hy4-preview weights are there. So for now, Hy4 is an API model with an open-weights promise, not a model you can download.

OpenRouter currently lists Hy4 at $0.417 per million input tokens and $1.2505 per million output tokens, which is a 50% discount on its listed rate.

Tencent Hy4 at a glance#

Hy4 pricing: Tencent vs OpenRouter#

Tencent quotes one price, OpenRouter shows another, and both have a discount running.

Tencent says the standard Hy4 price is the same as Hy4 Preview’s, even though the model improved. Tencent also made WorkBuddy and CodeBuddy free from November 4 to 15.

OpenRouter routes tencent/hy4 to a single provider, Tencent, and its endpoint data tags that deployment as fp4 quantization. There’s no fallback provider for this model on OpenRouter right now. If Tencent’s endpoint has a bad hour, your requests fail rather than move to another host. Our OpenRouter vs direct API comparison covers that trade-off for coding agents.

Calling Hy4 through OpenRouter#

OpenRouter’s API is OpenAI-compatible, so the model ID is the only Hy4-specific part of the request.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
    "model": "tencent/hy4",
    "messages": [{"role": "user", "content": "Summarize this stack trace and suggest a fix."}],
    "reasoning": {"effort": "low"}
  }'

Reasoning is on by default for Hy4 on OpenRouter, with “high” as the default effort and “low” and “none” as the other options. For short tool-calling loops, dropping the effort cuts output tokens, which is where most of the cost is. OpenRouter’s endpoint data also shows tool_choice accepts auto and none but not required, so a client that forces a tool call on every turn may need a config change.

If you run OpenCode, Hy4 slots in like any other OpenRouter model. The OpenCode custom providers guide shows where the model ID goes.

What’s new in Hy4#

Tencent lists these changes over Hy4 Preview:

• Visual understanding, for the first time. Tencent says Hy4 handles fine-grained image processing, chart analysis and mathematical visual reasoning.
• Computer use, for the first time. Hy4 can work with both command-line and graphical interfaces. Tencent says it had specialized training on professional software such as Blender and CAD.
• Long-horizon agent work. Tencent reports gains in instruction following, long-context comprehension and long agentic tasks.
• Game and front-end development. Tencent says Hy4 improved 19.1% on GameDev and 9.8% on GameLogic benchmarks over Hy4 Preview.
• Fewer hallucinations, by Tencent’s measure. In internal evaluations, the hallucination rate fell from 12.95% in Hy4 Preview to 7.2%, and errors from unsupported assumptions fell from 23.1% to 5.6%.

Hy4 benchmarks, with the caveats#

Most of Tencent’s numbers are its own evaluations. The headline one is a blind test run by the Tencent Hy team with Tencent Financial Technology: 213 financial professionals, including 153 experts from outside institutions, scored answers across 579 real-world financial tasks.

That test is narrow (finance work) and Tencent ran it. For a broader outside view, OpenRouter’s Hy4 page shows an Artificial Analysis Intelligence Index of 34.6, ranked 16th of 53 models there, and a Vals AI Terminal-Bench 4.0 score of 39.4%. OpenRouter says that index score beats 70% of the models it compares, though the top model listed there scores 59.1.

If you’re comparing open-weight options, Kimi K3.1 Pro already has downloadable weights and day-0 vLLM support, while Hy4’s weights are still pending.

Safety claims and what’s still missing#

Tencent says Hy4 was trained on defensive security data, with offensive exploitation material explicitly excluded from the training set. Experts from Tencent’s Xuanwu Lab and Keen Security Lab assessed areas including cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation, and Tencent reports they found no significant concerns about Hy4’s offensive cyber capabilities in the areas evaluated.

Several things aren’t public yet:

On Hy4 Flash, Tencent’s announcement is inconsistent. A section heading says “Hy4 Flash Recently Released”, but the text under it says Hy4 Flash “is scheduled for release soon”. Treat it as not yet available until it shows up on TokenHub or OpenRouter.

Is Hy4 worth testing?#

Hy4 fits a trial if you run cheap long-context agents and want image input and tool calls in the same model. At the current OpenRouter price, it costs less than half of Tencent’s standard rate. The catches are a single provider on OpenRouter, a discount with an end date, mostly vendor-run benchmarks, and weights that Tencent has promised but not shipped. Check Tencent’s Hugging Face page before you plan any self-hosting around it.