Models

Supported AI models

Onysoft AI Gateway offers models from the world's leading AI providers through a single API. Text models are used through the OpenAI-compatible /v1/chat/completions endpoint; image, video, audio and music models are used through the asynchronous generation endpoints.

Available Models

View all models and current prices →

Free Türkiye Models (:tr)

Models whose id ends with ":tr" are hosted in Türkiye and are free: no amount is deducted from your balance for these requests. They use the same API key, the same endpoints and the same request body; only the model id differs.

Limits

• Up to 100 requests per account (all your API keys combined) in any 24-hour period; if a custom limit has been set for your account, that limit applies. Free models are available only to verified accounts: signed in with Google or GitHub, or with at least one real payment (card/bank transfer).
• Chat requests can send up to about 64,000 input tokens and a 20 MB body per request, even when the model's context window is larger. Larger requests are rejected with a clear error.
• At most 4 requests are processed at the same time (per account for personal keys, per key for project and partner keys; if a custom value is defined for your account, that applies). Extra requests wait in a queue for a short time; if no slot frees up, 429 (tr_yogun or tr_concurrency_limit) is returned with Retry-After. Such a rejection is not deducted from your daily quota.
• Because they are free they appear in your usage log with an amount of 0.
• Models that accept image input: deepseek/deepseek-v4.1-flash:tr, qwen/qwen3.8-flash-next:tr, google/gemma-4-31b:tr, qwen/qwen3-vl-30b:tr, xiaomi/mimo-v2.6-pro:tr. Images sent to the other models are rejected with a clear error, without counting against your daily allowance.
• Some models think before answering: the reasoning text comes back in the reasoning field of the response and shares the max_tokens budget with the answer. For thinking models (DeepSeek, Qwen, Gemma, GLM and MiMo) set max_tokens to at least 2048, ideally 4096; with a very small max_tokens (e.g. 24) the whole budget can go to reasoning: content then comes back as null and finish_reason is "length". To turn thinking off and get faster answers, send "reasoning_effort": "none" to deepseek/deepseek-v4-flash:tr, deepseek/deepseek-v4.1-flash:tr, qwen/qwen3.8-flash-next:tr, xiaomi/mimo-v2.6-pro:tr. For google/gemma-4-31b:tr, use "chat_template_kwargs": {"enable_thinking": false} instead. google/gemma-4-31b:tr, qwen/qwen3-vl-30b:tr do not support reasoning_effort; if sent, the field is ignored. On xiaomi/mimo-v2.6-pro:tr, reasoning_effort accepts only: none, minimal, low, medium, high, xhigh; any other value is rejected without counting against your daily allowance.
• For speech to text, response_format can only be json or text; send recordings of about 30 seconds and files of at most 10 MB.

curl https://api.onysoft.com/v1/chat/completions \
  -H "Authorization: Bearer sk-ony-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v4-flash:tr",
    "messages": [{"role": "user", "content": "Merhaba"}]
  }'

Being free, these models can be busy depending on capacity. For production work that must not be interrupted, prefer the paid models in the catalog.

Typed Decision Endpoint (/v1/decisions)

This endpoint does not produce text. It returns a calibrated probability between 0 and 1 for the binary question you ask — "does this customer want a human?", "is this message reporting a blocking outage?". It is faster and cheaper than forcing a JSON schema on a chat model and parsing the answer, and parse errors do not exist here, because what comes back is already a number.

For detailed usage, question types and error codes, see the Typed Decisions API section.