temperature, top_p and image allowances follow DeepSeek’s API. These do not:
- Context is 1,048,512 tokens, 64 fewer than DeepSeek’s window.
max_completion_tokensis read as the output cap whenmax_tokensis absent (max_tokenswins when both are sent).response_format: {"type": "json_object"}always returns a JSON object. DeepSeek’s API can return any JSON value.- Image
detail: "low"is accepted only up to 512 px on the longer side. A larger image is refused with a 400 (DeepSeek’s API downscales it). - Images by URL: 60 seconds an image and 600 a request, as on DeepSeek’s API, but all of a request’s images within 300 seconds and about 36 MiB in total, at most 3 redirects, public hosts only.
- Not offered: the Files API (
file_id,file_data), web search, URL reading, the application APIs DeepSeek released on 2026-11-13, and FIM completion. - Tool call ids, log probability fields and stream framing have DeepSeek’s shapes but are made by OS3.
- A message that contains the literal text
<|deepseek_image|>next to a real image is refused with a 400. - Where the model is unsure of the next token, the text can differ from DeepSeek’s API or from another host of the same checkpoint.
temperature: 0does not give the same text every time.