Open-weight models through an OpenAI-compatible API, served on GPUs, racks and networks owned and operated by team.blue.
Sovereignty is a supply-chain property, not a badge. Every request is served by GPUs team.blue bought, in a facility team.blue operates, in a region you can name. There is no third-party inference provider in the path and no cross-border hop in the middle of a completion. If a customer asks where their prompt went, the answer is a rack — not a region code on someone else's status page.
Standard OpenAI-compatible routes, so existing SDKs, gateways and libraries keep working unchanged.
from openai import OpenAI client = OpenAI( base_url="https://llmfactory.ai/v1", api_key="sk-...", ) r = client.chat.completions.create( model="glm-5.3", messages=[ {"role": "system", "content": "You triage hosting support tickets."}, {"role": "user", "content": "My DNS change is still not live after 2 hours."}, ], stream=True, )
Models are allocated to your team and appear in the catalogue with their context window already set. Long conversations are summarised rather than truncated.
Scoped API keys per team with a request quota and a rate limit, so one runaway script returns 429 instead of taking capacity from everyone else.
Tokens, cost and latency per key and per model, daily. Enough to see a regression the day it appears and to bill it back internally.
Requests are placed on the tier that fits the model, so a small embedding job doesn't occupy a card meant for a 70B.
Embeddings, transcription and small models. Cheap tokens.
The everyday workhorse behind most production traffic.
Larger models, longer context, higher concurrency per card.
The newest tier, kept for the heaviest serving jobs.
Models are served with vLLM and Ollama on our own cards. Capacity is added tier by tier as demand justifies it, in the same region — not by renting somewhere else.
Access is restricted to internal use cases — product features, internal tooling and brand experiments running on our own infrastructure.
Public endpoints are planned: the same fleet, the same region, opened beyond internal workloads. No date promised yet.
Your local team.blue brand handles onboarding, model access and current capacity. They know what is available this quarter.
DK9 is where the fleet runs today. A second region is being built out now — same owned hardware, same API, one more place to pin your workloads. We will name it when the racks are powered, not before.