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 examples show how the model can be used for real tasks, but wider claims about its performance come from the companies themselves.
Listen to this story
The audio brief
Story brief
3 key pointsTogether AI made Tencent’s Hy4 available through its serverless chat-completions API on November 19, 2026. This gives developers a new way to use the model, rather than marking its first release. Hy4 can take images and text, return organized JSON data, use tools and handle a context window of up to one million tokens. Together has published prices for using the model. Its chart and game examples were run by the company and do not independently prove that the model works reliably in a wide range of situations.
- 01
Together lists prices of $0.83 for every million input tokens, $2.50 for every million output tokens and $0.042 for every million cached input tokens.
- 02
Tencent says errors caused by unsupported assumptions fell from 23.1% in Hy4 Preview to 5.6% in Hy4 in its own tests. This does not guarantee general accuracy.
- 03
Tencent reports an average score of 3.04 in a blind financial test involving 213 professionals and 579 tasks. However, the company says Hy4 still falls behind the best privately owned models on its hardest tests.
Developers can now send images and text to Tencent’s Hy4 through Together AI. This gives them access to a hosted model designed to read charts and work with software. Together announced that Hy4 was available through its serverless service on November 19, 2026. The model has a context window of one million tokens. Its price is $0.83 for every million input tokens and $2.50 for every million output tokens.
This gives developers a new way to use an existing model. It is not the first release of Hy4. Tencent announced Hy4 on November 4. It introduced the model as a system that can understand images and use computers. Together’s release makes these features available through its own chat-completions API. It also provides prices and examples that developers can try themselves.
Turning an image into a useful answer
Hy4 can take text and images, but it returns text. Its one-million-token context window is the amount of information it can work with at one time. This can include large groups of documents and long records of previous work. Together also says developers can adjust how much reasoning the model uses.
Two features help developers use the model’s answers in their applications. JSON output puts an answer into an organized data format instead of ordinary sentences. Tool calling lets developers give the model functions that it can use. In Together’s example, a developer puts an image in a message and asks the model to return a JSON object with the chart’s title and bars.
Using a computer involves more than reading an image. Tencent says it trained Hy4 to follow typed commands and work with software that people use through a visual screen. Examples include Blender and computer-aided design tools. Tencent product teams helped with its training and testing. The work covered office tasks, financial research and game development.
Together’s standard price for information sent to Hy4.
Information created by the model has a separate price.
Together lists a third price for cached input.
Rebuilding a chart and making a game in one file
Together’s launch article shows how Hy4 can turn an image into code. The example uses a screenshot of one of Together’s own charts. The company says Hy4 correctly read the chart’s title and all three values. A second request asked the model to recreate the chart in a single HTML file. This meant it had to do more than read the numbers: it also had to rebuild the chart’s appearance.
Together says the recreated chart kept the labels, values, colors, dashed line showing the upper limit and footnote. The font was different because the instructions asked for general fonts. This example shows how the model handled one particular chart-reading and rebuilding task. It was run by the company, not an independent test of how well the model handles unfamiliar charts.
For another task that used only text, Together asked Hy4 to make a new browser game in one HTML file, without libraries or outside files. Hy4 created a color-matching game called Polarity. Together says it opened the code without making changes, played a round, reached the end of the game and successfully restarted it. The model also added a title screen and a best-score record that stayed between rounds, even though these features were not requested.
Wider claims about reliability come from Tencent
Tencent’s earlier announcement gives a wider view of how the company says Hy4 performs at work. It reports that errors caused by unsupported assumptions fell from 23.1% in Hy4 Preview to 5.6% in Hy4. These numbers cover this one type of error in Tencent’s own tests. They do not mean that the model will always give correct answers.
Tencent also reports an average score of 3.04 in a blind financial test. The test involved 213 professionals and 579 tasks based on real-world work. Together’s model information includes an important warning alongside these reported improvements: Tencent says Hy4 still performs worse than the best privately owned models on its hardest internal tests. These include advanced mathematics and coding tasks that need many steps over a long period.
Developers can use the hosted model with the name tencent/Hy4 and a Together API key. The launch article includes a Python example for reading charts. Developers can also try the model in Together’s online playground before writing code.
Sources
- together.aiTencent Hy4 is now available on Together AI
- tencent.comTencent launches Hy4, joining the ranks of the world's leading open-source AI models
- together.aiTencent Hy4 API: Prices, benchmarks and documentation | Together AI
