Tencent launches Hy4 with vision and computer-use upgrades
Tencent launched Hy4 on November 4, 2026, adding visual understanding and computer-use capabilities to its flagship model. The company says it plans to release downloadable weights and system cards in the coming days. For now, the documented access routes are hosted services. Tencent announced the rollout.
For developers, Tencent Cloud’s model catalog makes the immediate change clear. The final model uses the API identifier hy4 and adds image understanding. It retains the preview’s listed 1M-token context window, so this launch should not be described as a new jump from a small context window to one million tokens.
What changes from Hy4 Preview
The earlier Hy4 Preview arrived on August 28. Tencent already positioned that version for coding, office work and research, with a context window exceeding one million tokens. Those existing features provide the baseline for assessing the final release. A demonstration of long document handling alone would not establish what improved in November.
Tencent says the new version can reason about images and operate command-line and graphical interfaces. Its announcement highlights training involving Blender and CAD, alongside longer multi-step tasks. The company lists WorkBuddy, CodeBuddy, Yuanbao and ima among the products offering Hy4.
The useful distinction is between understanding a screenshot and completing work inside an application. A model might identify a button correctly but still mishandle a later step, an unexpected dialog or a changed file. Teams evaluating computer use should test the whole workflow and inspect the saved result, rather than score only the first action.
API capabilities and practical limits
TokenHub documents a maximum input of 960k tokens and maximum output of 64k tokens for hy4. It also lists reasoning, structured output, function calling and caching. These separate limits matter when budgeting a request. The context-window label should not be treated as permission to send one million input tokens or generate one million output tokens.
OpenRouter’s live model catalog separately lists tencent/hy4 with 1,048,576 tokens of context and a 65,536-token completion limit. It accepts text and images and returns text. The supported parameter list includes tools and structured outputs, with reasoning enabled by default at the high setting. Low and none are also listed reasoning efforts.
Treat these as provider-specific specifications. Confirm the chosen endpoint’s request format, permissions and limits before moving an existing integration. Image input also serves a different purpose from image generation. For an agent, start with a small reversible task, provide only the tools it needs, and compare its final artifact with an explicit expected result.
Hy4 pricing and the launch discount
Tencent Cloud’s Hy4 promotion charges 50% of catalog rates through December 31, 2026, at 23:59:59 Beijing time. Catalog pricing resumes afterward. The amounts below are per million tokens.
| Service and rate | Input | Output | Cache hit |
|---|---|---|---|
| TokenHub catalog | CNY 6 | CNY 18 | CNY 0.30 |
| TokenHub promotion | CNY 3 | CNY 9 | CNY 0.15 |
| OpenRouter, November 4 | US$0.417 | US$1.2505 | US$0.021 |
OpenRouter’s figures are the observed aggregator rates for tencent/hy4, rather than a currency conversion of Tencent’s catalog. Recheck the chosen provider before deployment because prices can change.
As a simple calculation, 100,000 uncached input tokens plus 10,000 output tokens would cost CNY 0.39 under the TokenHub promotional rates, compared with CNY 0.78 at catalog rates. This illustrates token charges only. It assumes the stated quantities and does not estimate how many tokens a real task will consume.
For a useful cost comparison, keep the task, success criteria and tool access constant. Record total billed usage, retries and elapsed time. A cheaper request can become a more expensive completed job if it needs repeated correction. Cache-hit pricing should be considered separately from fresh input rather than applied to every prompt.
Performance claims still need outside testing
Tencent reports stronger results on workplace, game-development and research evaluations. Those results come from the company or its evaluation partners. They are vendor-reported evidence, not ByteForward testing or an independently established ranking.
A practical evaluation should include tasks where success is easy to verify and failure is costly enough to notice. For document work, check whether the answer preserves conflicting details and follows the requested format. For coding, run the relevant tests. For visual work, verify small labels and relationships instead of accepting a plausible description. Keep human approval around consequential actions while measuring whether the model finishes reliably.
What remains to be released
Tencent also describes Hy4 Flash as forthcoming. Final weights and system cards remain promised releases, rather than artifacts verified for this report.
Developers can begin a hosted evaluation now. Anyone planning self-hosting should wait for the actual final model artifacts, license and deployment requirements, then check them directly. Preview documentation can help explain the model family’s history, but it cannot settle the final release’s permissions or hardware needs.
For more on inspecting released model artifacts, see ByteForward’s report on Tencent’s DSpine code and checkpoint release.
