Model Management

The provider exposes LM Studio's native model-management endpoints, so you can control what is in memory without shelling out to lms:

provider = RubyLLM::Provider.resolve!(:lms).new(RubyLLM.config)

provider.native_models   # every downloaded model, with architecture,
                         # quantization, capabilities, loaded_instances, ...
provider.loaded_models   # just the ones resident in memory

instance = provider.load_model('qwen/qwen3-4b', context_length: 8192)
# => { "instance_id" => "qwen/qwen3-4b", "status" => "loaded", ... }
provider.unload_model(instance['instance_id'])

provider.download_model('qwen/qwen3-4b')  # fetch a new model onto disk

Method	Endpoint	Returns
native_models	GET /api/v1/models	every downloaded model as LM Studio describes it, as an Array of Hashes
loaded_models	GET /api/v1/models	the subset with at least one loaded instance
load_model(model, context_length: nil, **options)	POST /api/v1/models/load	the server's response Hash, including instance_id
unload_model(instance_id)	POST /api/v1/models/unload	the server's response Hash
download_model(model, **options)	POST /api/v1/models/download	the server's response Hash

Extra keyword options go into the request as-is; the server rejects keys it doesn't know.

Instances

load_model returns the instance_id that unload_model takes. Loading a model that is already resident starts a second instance (...:2), so hold on to the id you were given.

For the same reason these calls are sent non-idempotent: RubyLLM retries a POST that times out, and a retried load of a large model would hold the weights in memory twice.

The same thing can happen behind your back with decision models: a model you loaded yourself, then requested on demand by a judgment, can end up with two instances. See Decision Models.

TTL

ttl is not a load option on the native API. Set idle TTL per request through the OpenAI endpoints (the ttl request extra) or with lms load --ttl.

Example 14_model_management.rb lists, loads and unloads a model.