Models & Auto

Use the compute you own, then the remote reasoning you authorize. Model selection and permission to execute are separate decisions.

Auto is recommended

Auto evaluates the request type, conversation size, available local models, collaborator grants and presence, configured cloud access, and station routing policy. Short conversational tasks generally prefer local compute. Analysis and longer context may use a larger owned model, a collaborator, or your configured cloud reasoning route.

This is a routing heuristic, not a benchmark or a guarantee of the cheapest or best answer. Model size is an estimate of capacity. Auto can select remote reasoning immediately for a complex task; it does not always run a local model first. Review the route information returned with the request.

Choose a route deliberately

Local model: reasoning runs on your station's configured model host. A model host configured on another computer is a remote destination even if it uses Ollama.
Your cloud key: the provider processes the context sent for reasoning and bills your account. The documented station v1.5.24 Chat picker exposes the Anthropic BYOK route. Keys configured for another product surface are not automatically available in this picker.
Collaborator: reasoning runs on a teammate's station under a share grant. That host receives the permitted context and must be online with a model ready.
RailCall Assistant: an optional hosted service with its own orchestration, capacity, and data disclosure. Its external models use service-funded access, independently of your station's keys.

An explicit model choice pins that station request to the selected route. If it is unavailable, resolve the issue or deliberately select another route. Do not assume an unavailable model has been substituted successfully.

Set up local reasoning

With the station running, use another terminal to inspect the model advisor and installed models. Ollama must be installed and running to pull or serve a local model.

shell

railcall models suggest
railcall models list
# Replace TAG with a model tag suitable for your machine:
railcall models pull TAG

Advisor speeds are estimates. Measure response time and quality with your own tasks before choosing a large model. Memory pressure and cold loading can make a larger model slower without making it useful for your workload.

Keep reasoning on your device

In Studio Settings, enable the station routing policy's Local-only option when remote model calls are not allowed. Ensure the configured model host is on this computer. With no eligible local model, model work is blocked. This policy governs reasoning routes; it does not disable installers, telemetry, web tools, or integration networking.

Cloud and collaborator reasoning are subject to the station's data-egress policy. Configuring a key or share grant makes a route available to Auto; it does not mean every reasoning call opens a separate approval card. The action airlock governs execution. See Privacy & data boundaries.

Use a collaborator's model

Once both stations belong to the same team, the holder can create a model grant. Keep the holder's station and Ollama running. The member can then ask the shared model and inspect measured usage.

shell

# Holder: replace MEMBER and TAG with your teammate and installed model.
railcall team share-model MEMBER --model=TAG --tokens-per-day=20000

# Member:
railcall team reason "Compare a staged rollout with a full rollout" --max-tokens=800
railcall team compute

# Holder: revoke future use when needed.
railcall team unshare GRANT_ID

See Teams for membership setup. Grant limits do not turn shared hardware into an unlimited service. An offline host or exhausted allowance can prevent completion.

Understand the cost

Local and shared models use hardware, electricity, and operator capacity. BYOK requests incur your provider's charges. Hosted Assistant limits and funding are separate. Use measured token usage and the provider's bill to evaluate savings; there is no fixed percentage saving promised.

Patrik's deployment demonstrates a shared orchestration host. You can use existing local and team model capabilities today. A turnkey personal copy of the hosted optimizer is not documented as a shipped installation flow.