Together AI adds Tencent Hy4 with image input and a million-token context window
The hosted release includes tool calling and structured output. Together’s chart and game demonstrations offer a practical look, while broader performance claims remain company-reported.
Together AI’s November 19, 2026, release makes Tencent’s Hy4 available through its serverless chat-completions API; it is a new hosting option, not the model’s debut. Hy4 can take images and text, return structured JSON, call tools, and use a context window of up to one million tokens. For developers, the launch pairs those capabilities with published pay-as-you-go rates, while Together’s chart and game examples remain provider-run demonstrations rather than independent evidence of broad reliability.
01
Together lists Hy4 at $0.83 per million input tokens, $2.50 per million output tokens, and $0.042 per million cached input tokens.
02
Tencent says unsupported-assumption errors fell from 23.1% in Hy4 Preview to 5.6% in Hy4 in its internal evaluations; this is not a general accuracy guarantee.
03
Tencent reports a 3.04 average score in a blind financial evaluation of 213 professionals and 579 tasks, while noting Hy4 trails top proprietary models on its hardest tests.
Developers can now send images and text to Tencent’s Hy4 through Together AI, giving them hosted access to a model designed to read charts and work with software. Together announced serverless availability on November 19, 2026, with a million-token context window and rates of $0.83 per million input tokens and $2.50 for output.
This is a new route to an existing model, not Hy4’s original debut. Tencent announced Hy4 on November 4, introducing visual understanding and computer use to the family. Together’s launch brings those capabilities to its own chat-completions API, with published prices and examples developers can reproduce.
From an image to a usable response
Hy4 accepts text and images but returns text. Its million-token context window is the space available for the material supplied to the model, including large document sets and extended working histories. Together also lists adjustable reasoning effort, allowing developers to vary the model’s reasoning setting.
Two supported features help connect those responses to an application. JSON output puts an answer into a structured data format rather than free-form prose. Tool calling lets developers supply functions for the model to use. Together’s example sends an image inside the user’s message and requests a JSON object containing the chart’s title and bars.
Computer use goes beyond reading an image. Tencent says it trained Hy4 to work through typed command-line instructions and graphical software interfaces, including Blender and computer-aided design tools. Its training and evaluation involved Tencent product teams and workflows in office work, financial research and game development.
Together’s Hy4 rates
$0.83Input
Together’s standard price for input supplied to Hy4.
$2.50Output
Generated output has a separate rate from input.
$0.042Cached input
Together lists cached input as a third pricing category.
Editorial illustration for Together AI adds Tencent Hy4 with image input and a million-token context window.
A chart reconstruction and a one-file game
Together’s launch post illustrates the image-to-code path with a screenshot of one of its own charts. The company says Hy4 extracted the title and all three plotted values correctly. A second request asked for a single HTML file reproducing the chart, shifting the task from reading numbers to rebuilding a visual layout.
According to Together, the reconstruction preserved the labels, values, colors, dashed ceiling line and footnote. The typeface differed because the prompt requested generic fonts. That example demonstrates a specific chart-reading and reconstruction task; it is a provider-run demonstration, not an independent assessment of performance across unfamiliar charts.
For a separate text-only task, Together requested an original browser game in one HTML file, without libraries or external assets. Hy4 returned Polarity, a color-matching game. Together says it opened the unedited output, played a round, reached the game-over screen and restarted successfully. The model also added an unrequested title screen and a best score retained between rounds.
Broader reliability claims come from Tencent
Tencent’s earlier launch announcement offers a broader, company-run view of workplace performance. It reports that errors caused by unsupported assumptions fell from 23.1% in Hy4 Preview to 5.6% in Hy4. Those figures describe that error category in Tencent’s internal evaluations, not a general guarantee that answers will be correct.
The company also reports a 3.04 average score in a blind financial evaluation involving 213 professionals and 579 real-world tasks. Together’s model card carries an important qualification alongside the claimed gains: Tencent says Hy4 still trails top proprietary models on its hardest internal tests, including research-level mathematics and coding tasks that require many steps over time.
Developers can access the hosted model using tencent/Hy4 and a Together API key. The launch post includes a Python chart-reading example, while Together’s playground offers a way to try the model before writing code.
Reader comments
Newest comments first. Replies stay oldest first.