Find free models

Discover local zero-cost runtimes and provider free-tier models, and reorder Arka’s LLM fallback chain with the ARKA_MODEL_POLICY variable.

arka free_models
arka free_models --json
arka free_models --offline

Results distinguish local zero-cost runtimes, live OpenRouter zero-price models, and provider free-tier-eligible models. Free-tier availability still depends on account, region, quotas, and provider policy; Arka never labels a paid model free merely because it has a free quota. Use arka hybrid status to see which discovered local or hosted options are actually reachable.

Model policy

ARKA_MODEL_POLICY reorders the fallback chain used by arka hybrid and by the shared LLM fallback engine. It does not add or remove providers; it only changes which side of the chain runs first (or restricts it to one side).

Value	Behavior
local-first (default)	Try local runtimes (Ollama, vLLM, LM Studio, Exo) before hosted providers.
hosted-first	Try hosted providers before local runtimes.
local-only	Skip hosted providers entirely. Useful for offline or private runs.
hosted-only	Skip local runtimes entirely. Useful when local models are too slow or unavailable.

export ARKA_MODEL_POLICY=hosted-first
arka hybrid status

# Or set it for one command only
ARKA_MODEL_POLICY=local-only arka chat "summarize this"

The policy runs after your existing fallback config: LLM_AUTO_FALLBACK, AI_PREFERRED_PROVIDER, AI_PREFERRED_MODEL, and per-skill routing in llm-skill-models.json still choose the initial candidate list. The policy then filters and reorders that list. Unknown values fall back to local-first.