theLAB runs a mix of local and Ollama Cloud models. The goal is to keep everyday inference on hardware we own (cheap, private, always-on) and reach for the cloud only when a request genuinely needs a frontier-scale model.
The two tiers
| Tier | Where | Examples | Use for |
|---|---|---|---|
| Local | your-llm-host (Ollama over Tailscale) |
Kimi K2.7, mid-size coders | Default fleet AI, chat, tool-calling, summarization |
| Cloud | Ollama Cloud (*:cloud tags) |
glm-5.2, large frontier models | Distillation teachers, hard reasoning, big context |
your-llm-host is the fleet's local-LLM host. Prefer it over hosted APIs for routine work — it is free to run and keeps data on the tailnet.
Routing rules of thumb
- Default to local. Route to
your-llm-hostfirst; only escalate when quality is insufficient. - Cloud is metered. Cloud tags like
glm-5.2:cloudburn Ollama Cloud Pro limits — reserve them for teacher runs (see Knowledge Distillation Lab (theLAB Genesis)) and one-off hard problems. - Keep clients pointed at hostnames, not IPs. Hardcoded tailnet IPs break when a host moves; use the Tailscale name.
Accessing your-llm-host
Ollama listens on the tailnet. From any fleet node:
OLLAMA_HOST=http://your-llm-host:11434 ollama list
OLLAMA_HOST=http://your-llm-host:11434 ollama run kimi-k2.7 "hello"Gotchas
Bind address. For a remote host to serve other nodes, Ollama must listen on the tailnet interface, not just localhost. Set
OLLAMA_HOST=0.0.0.0:11434in the service environment (and rely on Tailscale ACLs for access control).
Don't mix up cloud and local tags. A
:cloudsuffix routes to Ollama Cloud even if the base model exists locally — easy way to silently spend quota.
See also: Jetson Orin as an Ollama Host, Serving Multiple Models on a Dual-GPU Box.