Ollama-compatible API

superfluid serves a subset of Ollama's API (inference, discovery and embeddings) on its HTTP listener, for any runtime. It does not read Ollama's model store, pull models or interpret Modelfiles.

Point an Ollama client at the server root:

superfluid serve unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_M --port 11434
curl http://127.0.0.1:11434/api/tags

Note

Use a different port if Ollama itself is listening on 11434.

from ollama import Client

client = Client(host="http://127.0.0.1:11434")
model = client.list().models[0].model
for part in client.chat(model=model, messages=[{"role": "user", "content": "Hello"}], stream=True):
    print(part.message.content, end="", flush=True)

superfluid launch ollama runs the ollama command line against the server; see Ollama CLI.

Model names are superfluid's model ids; id:latest is accepted. Load models at startup or through POST /v1/models/load; arbitrary paths and remote names are refused.

Routes

method	path	behaviour
GET	/	superfluid is running
GET	/api/version	{"version", "server": "superfluid", "compatibility": "ollama-inference"}
GET	/api/tags	every known model id
GET	/api/ps	loaded models and their context length
POST	/api/show	model info; for a loaded model, capabilities and chat template
POST	/api/chat	chat with history, tools, thinking, images, JSON output
POST	/api/generate	templated prompt; raw for plain completion; suffix for fill-in-the-middle
POST	/api/embed	normalized embeddings (embedding models only)
POST	/api/embeddings	legacy single-prompt form, unnormalized
any	/api/pull, /api/push, /api/create, /api/copy, /api/delete, /api/blobs/*	501

Discovery routes never load a cold model.

Chat and generate

curl -N http://127.0.0.1:11434/api/chat -d '{
  "model": "unsloth/Qwen3.8-27B-GGUF",
  "messages": [{"role": "user", "content": "Explain prefix caching briefly."}],
  "options": {"num_predict": 128, "temperature": 0.2}
}'

curl http://127.0.0.1:11434/api/chat -d '{
  "model": "unsloth/Qwen3.8-27B-GGUF", "stream": false,
  "messages": [{"role": "user", "content": "Return a planet name as JSON."}],
  "format": {"type": "object", "properties": {"name": {"type": "string"}}, "required": ["name"]}
}'

field	notes
model	required
messages (chat)	system, user, assistant, tool; images (base64), thinking, tool_calls, tool_name, tool_call_id. Empty = preload
prompt, system, images, suffix, raw (generate)	empty prompt with nothing else = preload; an empty suffix is none
template (generate)	empty only (the model's own template); any other is a 400
tools	Ollama function tools
stream	default true (NDJSON)
format	"json" or a JSON schema
think	true/false maps to enable_thinking; minimal, low, medium, high to reasoning_effort
logprobs, top_logprobs	chat only; top_logprobs 0-20
options	see below; null is none
keep_alive	any non-null value is a 400

Any other non-null field is a 400 unsupported parameter: <name>. raw and suffix requests refuse system, images, format, think and options.stop.

Options

option	constraint
num_predict	positive, or -1/-2 to fill the context
num_ctx	must equal the server's context window
temperature	≥ 0
top_p, min_p	0-1
top_k	≥ 0
seed	≥ 0, or -1 for random
repeat_penalty	> 0
presence_penalty, frequency_penalty	-2 to 2
stop	up to four non-empty strings

Any other option (num_gpu, num_thread, mirostat, ...) is a 400 unsupported option: <name>.

Records

Each streamed line is one JSON object. Reasoning comes back separately as message.thinking (chat) or thinking (generate). The terminal record:

{"model":"Qwen3-4B","created_at":"...","message":{"role":"assistant","content":""},
 "done":true,"done_reason":"stop","total_duration":812345678,
 "prompt_eval_count":31,"prompt_eval_cached_count":16,"eval_count":42}

• done_reason is stop, length, or load for a preload.
• prompt_eval_cached_count (extension) is the prefix served from the cache.
• load_duration, prompt_eval_duration and eval_duration are not reported for generation.
• Tool calls arrive whole on one record with done: false, just before the terminal record; function.arguments is an object.
• Errors before the body keep their HTTP status with {"error": "message"}; an error mid-stream is one {"error": "..."} line.

Embeddings

curl http://127.0.0.1:11434/api/embed -d '{"model": "nomic-embed", "input": ["first", "second"]}'

field	default	notes
input		string or array; empty = preload
truncate	true	false refuses oversized input
dimensions	the model's	truncate and re-normalize

llamacpp and mlx serve no embeddings.

Not supported

• keep_alive (residency is the server's: --idle-timeout and the model routes)
• model distribution and Modelfiles (501)
• Ollama metadata and phase timings: /api/tags lists a digest of the model id, size 0 and empty details; /api/show ignores system, template and options
• the fleet head does not serve these routes