Choose, run, override, and evaluate models safely.

Local-First Inference

The default model path keeps inference on the user's machine. The Go binary stays CGO-free; llama.cpp runs as a supervised subprocess with model files stored outside the repo. Local model artifacts are large, so MARS downloads and verifies them under the user cache instead of committing them. For the full filesystem ownership map, use the Files And State Reference.

Local Bundles And Performance Profiles

A local bundle maps the fast, reasoning, and coding tiers to model artifacts that fit the detected machine. The performance profile controls how aggressive that choice should be.

mars models eligible
mars models eligible --json
mars models list --eligible --json

mars models list
mars models list --provider registry --json

Setup And Download

mars setup prepares local inference. It creates the MARS user directory, writes config, detects hardware, installs llama-server artifacts, downloads model files when requested, and verifies the active local bundle before runtime.

make install
mars setup
mars doctor

mars setup \
  --inference local \
  --local-bundle auto \
  --download \
  --yes \
  --json

mars setup --inference defer --skip-download
mars run engineer --repo ~/my-project --dry-run

Routing Precedence

Role manifests declare intent with model: fast, model: reasoning, or model: coding. Runtime routing uses explicit overrides first, then manifest tiers, then role-name defaults, then a coding fallback. This prevents planning roles from accidentally using the heaviest coding tier when the manifest requested a lighter tier.

1. Repo model override

   .harness/model-overrides.yaml can override one tier or one role.

2. Manifest tier

   The role's model field wins when it names fast, reasoning, or coding.

3. Role-name default

   Known role names can route to an established tier when no explicit tier is present.

4. Coding fallback

   If nothing else matches, MARS falls back to the coding tier.

Repo-Owned Overrides

Overrides are deliberate target-repo configuration. They are useful when a project needs a different model for one tier, one role, or a remote provider route. They should be committed and reviewed like other harness changes.

mars models override \
  --repo ~/my-project \
  --tier coding \
  --provider ollama \
  --model qwen3.6:27b \
  --reason "Evaluate local coding candidate"

mars models override \
  --repo ~/my-project \
  --role engineer \
  --provider openai-compatible \
  --endpoint http://127.0.0.1:8088/v1 \
  --model repo-coder \
  --reason "Use dedicated coding server"

mars models override \
  --repo ~/my-project \
  --tier reasoning \
  --routing cloud \
  --provider anthropic \
  --model claude-sonnet-family \
  --api-key-env ANTHROPIC_API_KEY \
  --reason "Temporary high-quality review route"

Cloud Providers And OpenAI-Compatible Routes

Cloud routing is explicit. MARS stores provider metadata and api_key_env names in repo-owned config; it does not accept raw provider keys in flags, committed files, logs, traces, telemetry, JSON output, or errors. Enabling a cloud route sends selected assembled context, model messages, tool schemas, tool arguments, and tool results to that provider under its terms.

Credentials

Credentials are machine-local. A target repo can document which env names it expects, but secret values stay in the user environment or an ignored local file. The Authentication And Credentials Reference maps provider keys, private release auth, and integration secrets.

export ANTHROPIC_API_KEY=...
mars models evaluate \
  --provider anthropic \
  --model claude-sonnet-family \
  --api-key-env ANTHROPIC_API_KEY

mars models credentials write-local-env \
  --repo ~/my-project \
  --api-key-env ANTHROPIC_API_KEY \
  --yes \
  --json

Ollama

Ollama is a broad catalog and explicit swap provider, not an automatic default-promotion path. Operators can list local Ollama models, evaluate any model by name, and configure an explicit tier or role override when they accept responsibility for that route.

mars models list --provider ollama
mars models list --provider ollama --json

mars models evaluate \
  --provider ollama \
  --model qwen3.6:27b \
  --json

mars models override \
  --repo ~/my-project \
  --tier coding \
  --provider ollama \
  --model qwen3.6:27b

Model Evaluation

MARS treats new models as candidates until they pass harness-specific benchmark evidence. Newer or larger models do not become defaults because of model-card claims. A promotion must preserve tool-call reliability, structured JSON behavior, repo-ticket reasoning, speed, memory practicality, immutable artifact metadata, and checksum reproducibility.

mars models evaluate
mars models evaluate --json

mars models evaluate \
  --endpoint http://127.0.0.1:8088/v1 \
  --model repo-coder \
  --provider openai-compatible \
  --repo ~/my-project \
  --save-report \
  --json

mars models evaluate \
  --endpoint http://127.0.0.1:8088/v1 \
  --model repo-coder \
  --revision immutable-artifact-id \
  --sha256 SHA256_VALUE \
  --source https://example.invalid/model-card

Runtime Operations

Troubleshooting