Skip to content

Tencent Hy4 arrives on Together AI with image input

Short answer

Together AI now serves Tencent's flagship Hy4 model on its serverless inference platform. It is the first model in the Hy family to accept images with text, offers a 1M-token context window, tool calling and JSON output, and costs $0.83 per 1M input tokens and $2.50 per 1M output tokens.

What this means for operators

If your team re-keys numbers from screenshots, scanned reports or charts into a CRM or spreadsheet, a vision model that returns JSON through a standard API is worth a small trial. Pull ten real documents, send them with a fixed JSON schema and compare the output with what your people typed. Price the pilot from the listed token rates, with cached input at $0.042 per 1M tokens. Treat Together's tests as a vendor demo on clean inputs: your messy files decide whether it holds up.

Together AI has added Tencent's Hy4 to its Serverless Inference service. Developers call it with the model string tencent/Hy4 through the chat completions API.

What Hy4 offers

Hy4 is Tencent's flagship model and the first in its Hy family to take images as well as text. It returns text. Together lists a 1M-token context window, tool calling through the standard tools parameter, and JSON output enabled by setting the response format to a JSON object.

Tencent describes Hy4 as the successor to Hy4 Preview, adding image input and computer-use capabilities. According to Tencent, the model was trained to use specialized programs like Blender and CAD tools. The company also states that its own product teams trained and tested the model on tasks such as office productivity, financial analysis, and building games with Unreal Engine, Unity, and Godot. Those are Tencent's claims, relayed by Together; the post offers no benchmark figures.

Pricing

On Together's serverless tier:

  • $0.83 per 1M input tokens
  • $0.042 per 1M cached input tokens
  • $2.50 per 1M output tokens

What Together tested

Together ran two informal checks. In the first, it gave Hy4 a screenshot of one of its own bar charts and asked for the contents as JSON. The model returned the title and all three bar values as shown. Asked to rebuild the same chart as a single HTML file, it kept the layout, labels, values, colors, a dashed ceiling line and the footnote. Fonts differed because Together asked for generic ones.

In the second, a text-only prompt asked for a small browser game in one HTML file with no libraries. Hy4 produced a game called Polarity, in which players catch falling blocks matching their color and switch colors with the space bar. Together ran the file unchanged, played a round, reached the game-over screen and restarted. The model also added a title screen and a persistent best score without being asked.

These are single examples chosen and run by the host, not a systematic evaluation.

Trying it

The post includes a Python example that sends a base64-encoded image to the model and requests a JSON object back. It needs the Together SDK and an API key. Together also offers a playground for testing without code.

Source: Together AI · In the Atlas: Together AI →

Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.