ByteDance: UI-TARS 7B

UI-TARS 7B is a multimodal model from ByteDance, released on 22 July 2025 under the model id bytedance/ui-tars-1.5-7b. It belongs to the UI-TARS 1.5 family, a line of vision-language agents built for GUI-based environments. ByteDance describes it as optimized for desktop interfaces, web browsers, mobile systems and games, so it is best understood as a model designed around interacting with screens rather than a general purpose chat model. Its knowledge cutoff is 31 January 2025.

The model accepts images and text as input and returns text as output, which means it can look at a screenshot or interface capture alongside an instruction and produce a textual response. Its context window is 128,000 tokens, large enough to hold long interaction histories or detailed interface state alongside instructions, and its maximum output is 2,048 tokens per response. As a 7B parameter model it is relatively small for a vision-language agent, which matters when you are planning latency and throughput for agent loops that call the model repeatedly.

Those capabilities point to a specific set of practical uses. Because it takes screenshots as input and is tuned for GUI environments, it fits automation and agent workflows where the model needs to interpret what is on screen: reading a web page or application window, following a multi-step task across a desktop or mobile interface, or acting inside a game environment. The large context window helps when a task requires several screenshots or a running history, and the image input makes it possible to describe, summarize or reason about interface content rather than only plain text.

You call the model through api14's OpenAI-compatible API using the model id bytedance/ui-tars-1.5-7b. The supported request parameters are frequency_penalty, logit_bias, logprobs, max_tokens, presence_penalty, repetition_penalty, seed, stop, structured_outputs, temperature, top_k, top_logprobs and top_p. Since the interface is OpenAI-compatible, you can point an existing OpenAI client at api14 and keep your current request structure, adjusting the model id and any of the parameters above as needed. The seed parameter is useful for reproducible runs, structured_outputs helps when you want the response to follow a fixed shape, and the sampling controls give you fine-grained influence over how deterministic or varied the output is.

The main thing to weigh is the output limit. At 2,048 tokens per response, the model is suited to concise answers, action descriptions or short reasoning steps rather than long generated documents. If your task needs a lengthy written output, or if it does not involve images or interface state at all, a text-only model with a higher output ceiling may fit better. Within its intended role, however, UI-TARS 7B is a focused option for screen-aware agents, and the combination of image input, a 128,000 token context window and OpenAI-compatible access through api14 makes it straightforward to trial in an existing pipeline.