OpenAI & Anthropic
compatible inference

Point your existing SDK at this host, use your Lumid PAT as the API key, and call chat.completions or messages against a reasoning model — pick one with the model field (see /v1/models) or omit it for the default. Streaming on both shapes. Auth required — no PAT, no serving.

Get a token → Quick start Endpoints

196K

context window

~25/s

output tok/s

~85/s

aggregate (8 streams)

endpoints

01Get a token

Every request authenticates with a Lumid Personal Access Token (PAT). It's the only credential you need — the same token is your API key for every SDK and curl call below.

1 · Sign in	Go to lum.id and sign in (or create an account).
2 · Open Tokens	Head to lum.id/dashboard/tokens → Create token. Name it, set an expiry, and copy the value — it's shown once.
3 · Use it	Pass it as `Authorization: Bearer <pat>`, or as the `api_key` in any OpenAI / Anthropic SDK.

Treat a PAT like a password — it carries your identity and rate-limit tier. Store it in an env var, never commit it. Revoke and rotate from the same Tokens page if it leaks.

02Quick start

Model id is cyankiwi/MiniMax-M2.7-AWQ-4bit — or omit model and the server fills in the default.

Reasoning model — read this first. The model thinks before answering. The answer is in .reasoning (OpenAI shape) or the first thinking block (Anthropic shape) — .content is null / empty unless the model also emits a text reply. Set max_tokens ≥ 512 or the output is all reasoning with no answer.

OpenAI SDK

pip install openai

from openai import OpenAI

client = OpenAI(
    base_url="https://kv.run:5000/v1",
    api_key="<YOUR_LUMID_PAT>",
)
r = client.chat.completions.create(
    model="cyankiwi/MiniMax-M2.7-AWQ-4bit",
    messages=[{"role": "user",
               "content": "Hello"}],
    max_tokens=1024,
)
msg = r.choices[0].message
# answer is in reasoning; content may be null
print(msg.reasoning or msg.content)

Anthropic SDK

pip install anthropic

import anthropic

client = anthropic.Anthropic(
    base_url="https://kv.run:5000",
    api_key="<YOUR_LUMID_PAT>",
)
m = client.messages.create(
    model="cyankiwi/MiniMax-M2.7-AWQ-4bit",
    messages=[{"role": "user",
               "content": "Hello"}],
    max_tokens=1024,
)
# first block is thinking, last is the text reply
print(m.content[-1].text)

03Endpoints

All under /v1/. Add "stream": true for SSE on the generative endpoints.

GET /v1/models	List all deployed model ids.
POST /v1/chat/completions	OpenAI chat completion. Streaming.
POST /v1/completions	OpenAI text completion. Streaming.
POST /v1/embeddings	OpenAI embeddings (if the model has an embedding head).
POST /v1/messages	Anthropic messages. Streaming.
POST /v1/messages/count_tokens	Anthropic input-token count, no inference.

Available models. Default is cyankiwi/MiniMax-M2.7-AWQ-4bit (196K context). A 26B-parameter alternative is also available — pass its id from /v1/models in the model field to select it.

04Auth & limits

Auth	Required. Present a Lumid PAT as `Authorization: Bearer <pat>`. Anonymous requests get `401` — there is no free tier.
Rate limit	Per-principal request cap, keyed on your PAT subject — limits are per-identity, not per-IP.
Streaming	SSE pass-through. OpenAI `stream=True` and Anthropic `messages.stream()` both work unchanged.
Model field	Optional — omit it and the default is injected. Pass `cyankiwi/MiniMax-M2.7-AWQ-4bit` to be explicit.
Errors	`504` upstream timeout · `502` backend unreachable · `503` no backend configured.