Chat and coding

Chat and coding run on POST https://ai.ollalink.com/v1/chat/completions with your GPU API key (X-NH-GPU-Key), in the OpenAI request shape (model, messages, max_tokens, temperature, stream, tools, tool_choice, response_format). Both are always on, and since 7 November 2026 both lanes are served by the same coding-grade model with the same limits: use chat for everyday questions, explanations and instruction following, coding for code, bug fixes and agents. Replies name the lane you sent (chat or coding).

Setting	`chat` and `coding`
Default sampling	temperature 0.6, top-p 0.9, top-k 64, min-p 0.05, repeat penalty 1.1 (send your own to override)
Output	`max_tokens` 1–8,192 (default 8,192); `max_completion_tokens` means the same and has the same cap
Input	the whole request (messages, tools, formatting) plus `max_tokens` must fit the lane’s context (`400` with the token counts if not); body at most 1 MiB

Answers directly: no reasoning part by default; small tasks typically take 0.4–2.5 s. To think first on a hard task, send "chat_template_kwargs": {"enable_thinking": true}: the reasoning comes back in message.reasoning (streaming: delta.reasoning), the answer in message.content, and the reasoning counts towards max_tokens (expect roughly three times the tokens and time).
Tools (OpenAI format): tools with tool_calls replies, streamed or not, several calls in one reply; tool_choice auto, required, none or one named function; response_format json_object / json_schema. Your client runs the tools; ranges, patterns and formats in your schema are not enforced by the model, so check them in your tool.
Image input: add {"type": "image_url", "image_url": {"url": "data:image/png;base64,…"}} parts to a message — screenshots of terminal output, error dialogs, diagrams. Keep the body under 1 MiB (images under about 700 KB before base64).
Long answers: use "stream": true. A non-streaming request is dropped if nothing arrives for three minutes.
Conversations: send the same X-Conversation-Id header (up to 128 characters) on every turn of one conversation, so follow-up turns reuse the cached prompt and start faster.
When busy, a request waits briefly, then answers 503 with Retry-After; retry then.
Unfiltered: the model does not refuse requests; apply your own content rules if your product needs them.
Usage is counted in tokens (prompt plus answer, reasoning included) and shows under Usage in the dashboard.

curl https://ai.ollalink.com/v1/chat/completions -H "X-NH-GPU-Key: $YOUR_GPU_KEY" -H "Content-Type: application/json" \
  -d '{"model": "chat", "messages": [{"role": "user", "content": "Give me two tips for a job interview."}]}'

curl https://ai.ollalink.com/v1/chat/completions -H "X-NH-GPU-Key: $YOUR_GPU_KEY" -H "Content-Type: application/json" \
  -d '{"model": "coding", "stream": true,
       "messages": [{"role": "user", "content": "Write a Python function is_valid_ipv4(s) that rejects leading zeros."}]}'

Vision and OCR use the same route with an image part; see Models. For translation, use POST /v1/translate — it keeps numbers and placeholders intact.