Tencent Hy4 adds vision and computer use as final API launches

Tencent released the final version of Hy4 on November 4, adding image understanding and computer interaction. The company announced API access through Tencent Cloud and OpenRouter, alongside integration into tools including WorkBuddy and CodeBuddy. The change expands the material the model can work with beyond text alone.

Tencent Cloud distinguishes the new hy4 identifier from hy4-preview. Its model table gives both a one-million-token context window, a maximum input of 960,000 tokens and maximum output of 64,000. Image understanding is added to the final model’s row. Structured responses, function calling and context caching appear for both versions. Structured output makes responses easier for programs to parse; function calling invokes configured external tools. Caching reuses earlier context computations.

The platform’s API guide lists Hy4 among callable models and describes compatibility with OpenAI and Anthropic API formats. It also separates regional endpoints: Singapore has global resource scheduling, Guangzhou covers mainland China and Silicon Valley covers the US. Those are the document’s scheduling scopes for the connection routes.

Tencent describes analyzing complex charts and combining information from multiple images as use cases. Its computer-use description covers graphical interfaces and command-line interaction. Tencent presented these as company demonstrations.

OpenRouter lists the final model as tencent/hy4, accepting text and images and returning text. The listing also supports tool calling and structured output. It gives the context window as 1,048,576 tokens and the maximum completion as 65,536, more precise figures than the rounded limits in Tencent Cloud’s table.

A temporary discount on unchanged list rates

Tencent Cloud’s dollar schedule lists Hy4 and its preview at the same rates: $0.834 per million input tokens, $2.501 for output and $0.042 for cached input. A separate notice discounts final Hy4 usage by 50% through year-end. The promotion changes settlement during its stated window while retaining the published list rates. This report checked the published conditions; it did not make paid inference calls or reconcile an invoice.

OpenRouter’s Tencent Cloud provider row also displayed a 50% discount when checked on November 4. Its visible input and output prices were $0.417 and $1.251 per million tokens, with cached reads at $0.021. These are observed provider prices; billing follows the conditions of the route used.

For downloadable weights, the artifact name matters. Tencent’s public Hy4-preview model card describes a backbone with 770 billion parameters, of which 49 billion are active per token. The first figure describes overall scale, while the second covers the portion used for each token. Its table excludes an additional prediction layer. The card describes software, office-analysis, game-development and scientific uses. The preview weights carry the Apache 2.0 license, and the repository separately links the standard preview and an FP8 version.

Those files retain the preview name even as the final API identifier becomes available. During this check, OpenRouter’s link to a final-name Hugging Face repository returned an authorization error. API availability was verified, but public access to the final weight files at that address could not be established. The accessible preview repository therefore remains a separately identified artifact.

Tencent Cloud’s discount ends at 23:59:59 Beijing time on December 31, or 15:59:59 UTC. Its notice limits the discount to usage incurred within the promotional period, after which list pricing resumes.