Grimoire
What Lumen's documentation will never tell you.
What is this

Shadow is the cluster-local LLM service. No traffic leaves the switch. No usage is reported to Lumen. No rate limits shared with anyone outside this node. Use Chat for interactive sessions or the API for programmatic access.

The soul economy

Every new initiate is granted 20 souls, replenished at 0.05/hour. All cluster models are currently free. Pricing is listed on the Models page — shown for accounting purposes only.

API access

Get a key from your Profile. Drop-in OpenAI-compatible — swap the base URL and keep your existing code:

from openai import OpenAI

client = OpenAI(
    base_url="http://shadow.ncsa.illinois.edu/api",
    api_key="sk-shdw-your-key-here",   # not a Lumen key
)
resp = client.chat.completions.create(
    model="Qwen/Qwen2.5-7B-Instruct-AWQ",
    messages=[{"role": "user", "content": "Hello from the dark side."}],
)
print(resp.choices[0].message.content)
Available models
qwen3-coder-next Qwen3-Coder (next), served via the Lumen NCSA service.
inkling-small ThinkingMachines Inkling-Small (NVFP4, 2-node GB10 tensor-parallel). Large reasoning model with 1M context and DSpark speculative decoding. General-purpose chat and reasoning.
deepseek-ocr-2 Vision-specialized OCR model. Extracts text from images. Send images as base64 image_url content parts via /v1/chat/completions. Text-only prompts produce garbage — that is expected, not a bug.
qwen3-embedding-0.6b Text-to-vector embedding model (1024-dim). Uses /v1/embeddings, not chat. Vectors are semantically meaningful — cosine similarity reflects semantic closeness.
qwen3-reranker-0.6b Relevance-judgment model, NOT a chat model. Expects a yes/no relevance-judgment prompt (System + Instruct/Query/Document format). Returns 'yes' or 'no'; use logprobs for graded relevance scoring. Off-topic prompts → 'no' is correct behavior, not a bug.
API clients

API Clients let you create named organizations with separate API keys and usage tracking — useful if you want to attribute compute cost to a lab or project without sharing your personal key.