UI-TARS 1.5 7B API

Billed from your GPUniq balance as you use it — no subscription, no monthly minimum, no per-seat fee. Prices refresh from the live catalog hourly.

API identifier

ui-tars-1.5-7b

Pass this exact string as the "model" field of your request.

Type

Chat & text

Served by ByteDance.

Context window

128,000 tokens

About roughly 96,000 words — about 192 printed pages. Prompt, conversation history and attachments all count against it.

Max output

2,048 tokens

Ceiling for a single reply. Set max_tokens below it to cap cost per request.

Accepts

Images, Text

What you can put in the request body besides plain text.

Returns

Text

Tokenizer

Other

Determines how your text splits into billable tokens.

Available since

2025-07-22

Knowledge cutoff

2025-01-31

The model has no built-in knowledge of events after this date.

Upstream moderation

No

No vendor-side safety filter is applied on top of the model.

• Function calling: not supported

  Send tool definitions, get back the call the model wants made.

• Structured outputs: supported

  Replies constrained to your JSON Schema.

• Vision input: supported

  Accepts images in the message content.

• Extended reasoning: not supported

  Thinks before answering; thinking tokens bill as output.

curl https://api.gpuniq.com/v1/openai/chat/completions \
  -H "Authorization: Bearer YOUR_GPUNIQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ui-tars-1.5-7b",
    "messages": [{"role": "user", "content": "Explain rate limiting in one paragraph."}]
  }'

Create a key on /chat and send it as a bearer token. The same key works across every model in the catalog, so switching models means changing one string.

frequency_penalty

Discourages repeating tokens the model has already used often.

logit_bias

Nudges specific tokens up or down before sampling.

logprobs

Returns token probabilities — handy for confidence scoring and evals.

max_tokens

Hard cap on the reply length. Also your cost ceiling per request, since output is the expensive half of the bill.

presence_penalty

Pushes the model toward introducing new topics.

repetition_penalty

Blunt anti-loop control, mostly useful on open-weight models.

seed

Asks for repeatable sampling. Best effort on every vendor — it makes runs similar, it does not make them identical.

stop

Stop sequences that cut generation as soon as they appear.

structured_outputs

Schema-constrained decoding: the reply is guaranteed to parse against the JSON Schema you supply, so no retry loop around JSON.parse.

temperature

Randomness. Near 0 for extraction and classification, higher for drafting and ideation.

top_k

Limits sampling to the k most likely next tokens.

top_logprobs

Returns the n most likely alternatives per position.

top_p

Nucleus sampling — an alternative to temperature. Tune one or the other, not both.