API compatibility

This page lists what Spark's OpenAI and Anthropic compatible routes read and what they refuse. Anything not listed is ignored without an error.

OpenAI routes

Route	Supported
GET /v1/models	Yes. Lists ids and aliases, owned_by: "spark". On a central, the nodes' models too.
POST /v1/chat/completions	Yes. See below.
POST /v1/completions	Yes. prompt as a string or a one element array.
POST /v1/embeddings	Yes. For bert, nomic-bert and gemma-embedding2 models.
POST /v1/responses	No.
Audio, images, files, batches, fine-tuning, moderation	No.

Chat completions parameters

Parameter	Behavior
model	Required unless a default model exists. Matches an id, an alias or a file name, ignoring case.
messages	Required, non-empty. Roles system, developer (as system), user, assistant and tool. Fields read: content, name, tool_call_id, tool_calls, reasoning_content, reasoning.
Content parts	text and input_text, image_url and input_image. input_audio and file are refused with 400.
stream, stream_options.include_usage	Yes. Ends with data: [DONE].
max_completion_tokens, max_tokens	Yes. Must be positive.
temperature, top_p, top_k, min_p	Yes.
presence_penalty, frequency_penalty	Yes.
repetition_penalty, repeat_penalty	Yes, as extensions.
seed	Yes.
stop	Yes. A string or an array.
n	Only 1. Other values give 400.
tools	Yes. Function tools, in {"type": "function", "function": {...}} form or flat. Each needs a name.
tool_choice	Only "none" is read. Other values are ignored.
response_format	json_object and json_schema.
reasoning_effort, think, chat_template_kwargs.enable_thinking, reasoning.effort	Turn thinking on.
prompt_cache_key	Read as a cache key.
logprobs, top_logprobs	Ignored. logprobs is always null in the response.
user, metadata, store and others	Ignored.

Completions parameters

The sampling parameters above, plus prompt (required) and max_tokens (default 16). Token arrays and batches are refused. n must be 1.

Embeddings parameters

Parameter	Behavior
model	An embedding model. A chat model gives 400.
input	A string, or a non-empty array of strings. Token arrays are refused.
encoding_format	float (default) or base64. Anything else gives 400.
dimensions	Truncates and normalizes again. More than the model has gives 400.

On a central, a model that only a node has does not answer embeddings: 501.

Response fields

usage has prompt_tokens, completion_tokens, total_tokens and prompt_tokens_details.cached_tokens. Reasoning comes as reasoning_content. The x-request-id header names the request in the request log.

Anthropic routes

Route	Supported
POST /v1/messages	Yes. See below.
POST /v1/messages/count_tokens	Yes. Returns input_tokens.
Message batches, the Files API, the models list	No.

Messages parameters

Parameter	Behavior
model	As for OpenAI.
max_tokens	Required. 400 without it.
messages	Required, non-empty. content as a string or an array of blocks.
system	A string, or an array of text blocks.
Blocks	text, image (base64 or url source), thinking, redacted_thinking (ignored), tool_use, tool_result. document is refused with 400.
stream	Yes. Typed events.
temperature, top_p, top_k	Yes.
stop_sequences	Yes.
tools	Tools with an input_schema. Server tools without a schema are skipped.
tool_choice	Only {"type": "none"} is read.
thinking	type of enabled or adaptive turns thinking on.
cache_control	On blocks, on system, on tools, or at the top level. Writes a KV snapshot at the breakpoint.
metadata, service_tier, mcp_servers and others	Ignored.

Messages response fields

stop_reason is end_turn, max_tokens, stop_sequence or tool_use. usage has input_tokens, cache_read_input_tokens, cache_creation_input_tokens and output_tokens. A thinking block has an empty signature.

Authentication and errors

Case	OpenAI routes	Anthropic routes
Key sent as	Authorization: Bearer or x-api-key	x-api-key or Authorization: Bearer
No or wrong key	401, invalid_request_error	401, authentication_error
Unknown model	404, not_found_error	404, not_found_error
Invalid request	400, invalid_request_error	400, invalid_request_error
Generation failed	500, api_error	500, api_error

Known gaps

• No continuous batching. Requests to one model run one at a time.
• No n above 1 and no logprobs.
• No Responses API.
• No audio input, and no image or video generation.
• No per-key rate limits, quotas or model allow lists.
• No context shifting. A prompt that exceeds the context is refused.