Community provider alpha

Contribute one safe local route.

Check one local model first. A verified account can create a token; the node stays pending until review.

Keep control local. A node serves only models you select, uses an outbound authenticated connection after enrollment, and can leave at any time. Do not use personal OAuth subscriptions or unreviewed model sizes as Community capacity.

01 / inspect

Discover without inference

Read local Ollama tags and hardware data. The check does not load a model, write a credential, or contact Tekiz.

02 / choose

Choose safe local models

Start with the installed chat model that fits your measured resident-memory budget. You can add a second, approved model for a specific task later; one local model remains resident by default.

03 / enroll

Create a pending node

A verified account creates a short-lived token after the check. The node remains pending; activation and worker startup are separate review steps.

Private alpha

Need private-alpha access?

Request private-alpha access when account setup is unavailable or you need help reviewing a constrained host. This request does not enroll a node or grant access to customer work.

Start here

Run the no-inference capacity check.

Set the local tag you intend to offer, then run the check from the Tekiz checkout. It queries only local Ollama metadata and returns a JSON recommendation. It refuses cloud, embedding, OCR, unknown-size, and over-budget tags.

For a constrained host, keep one resident model and choose a smaller tag. Mixture-of-Experts tags are judged by total resident weights, not active parameters.

SAGE_ROUTER_COMMUNITY_MODELS=qwen2.5:7b-instruct \
scripts/enroll_community_node.sh --check

Optional specialization

Assign a local model to the work it suits.

After both models pass the capacity check, map a task to an enrolled local tag. This helps Fusion send implementation work to the smaller model while reserving the larger one for research.

The map selects the requested tag; it does not keep both models in memory. A constrained node still runs one job and one resident model at a time.

SAGE_ROUTER_COMMUNITY_MODELS=qwen2.5:7b-instruct,qwen2.5:14b-instruct \
SAGE_ROUTER_COMMUNITY_TASK_MODEL_MAP=implementation=qwen2.5:7b-instruct,research=qwen2.5:14b-instruct \
scripts/enroll_community_node.sh --check

Then, not before

Enroll and run the outbound worker.

Create the short-lived token in Account after the check. Enrollment creates a pending node; after activation, the worker polls for one bounded job at a time. This command is separate from the check.

# Enrollment writes an owner-only node credential and contacts the control plane.
SAGE_ROUTER_COMMUNITY_NODE_LABEL=my-node \
SAGE_ROUTER_COMMUNITY_MODELS=qwen2.5:7b-instruct,qwen2.5:14b-instruct \
SAGE_ROUTER_COMMUNITY_TASK_MODEL_MAP=implementation=qwen2.5:7b-instruct,research=qwen2.5:14b-instruct \
SAGE_ROUTER_COMMUNITY_ENROLLMENT_TOKEN=restricted-token \
scripts/enroll_community_node.sh

# Start only after the node is activated.
python3 scripts/run_community_node.py

Operating rules

Capacity is a promise you can keep.

01Start with one selected local chat model and one accepted Community job at a time. More parallelism requires measured VRAM, thermals, and latency headroom.
02Do not advertise cloud catalog models, embeddings, OCR, or a personal provider subscription. Community work must use the admitted local model list.
03Use --check again after changing hardware, Ollama tags, or residency settings. A successful check is not an instruction to run a larger model.
04Community Fusion requires customer opt-in. Prompts are never put on a settlement ledger, and a Community worker never returns directly to the customer.