Models

LM Studio serves whatever models you have downloaded — chat models, embedding models and decision models — so the provider asks the server what it has rather than relying on a fixed list.

provider = RubyLLM::Provider.resolve!(:lms).new(RubyLLM.config)
provider.list_models.each { |model| puts "#{model.id}  #{model.capabilities.join(', ')}" }

You don't need a catalog to use a model: because :lms is a local provider, RubyLLM assumes any id you pass exists and LM Studio loads it on demand.

What listing enrichment adds

list_models starts from the OpenAI-compatible /v1/models listing, which carries little beyond an id, and merges in LM Studio's native listing. It tries GET /api/v1/models first and falls back to GET /api/v0/models; when neither answers — an older build, or the native API turned off — the plain listing is still returned.

Model#name becomes the display name where one is reported ("Qwen3 27B" rather than "qwen/qwen3.8-27b"), and Model#metadata picks up whatever the server knows:

Key	From	Example
owned_by	the OpenAI listing	"organization_owner"
publisher, arch, quantization, compatibility_type, state	v0 and v1	"qwen", "qwen35", "Q4_K_M", "gguf", "loaded"
display_name, description	v1	"Qwen3 27B"
params_string, size_bytes, bits_per_weight	v1	"27B", 17742039110, 4
variants, selected_variant	v1	the downloaded quantizations, and which one is selected
reasoning_options	v1	[{type: "effort", values: ["none", "low", "medium", "xhigh"], default: "xhigh"}, {type: "toggle"}]
loaded_context_length	v1	8192 — the loaded instance's window, which can be smaller than context_window
remaining_ttl_seconds	v1	1050 — how long before an idle JIT-loaded instance evicts
decision_type	v1, decision models only	"lfm2-d1-omni", "kev", "clef"

Capabilities

Kind of model	capabilities	modalities
Chat	streaming, structured_output, plus function_calling and tool_choice (tool-trained), vision, reasoning as the listing reports	text in (plus image for vision), text out
Embedding	none	text in, embeddings out
Decision	judgment only	text in (plus image with a vision adapter), judgment out

streaming and structured_output hold for every chat model LM Studio serves: the server constrains generation with a grammar built from your schema, so structured output works even on weights with no tool training — whether the values are any good is a separate question.

Decision models don't chat, so they get the judgment capability and output modality and nothing else. See Decision Models.

Model lifecycle: JIT loading, TTL, auto-evict

The server manages memory rather than a fleet:

JIT loading

With just-in-time loading enabled (the server default), /v1/models lists every downloaded model and a request for an unloaded model loads it on demand. Expect a long time to first token on that first request. With JIT off, only loaded models are served.

TTL

JIT-loaded models default to a 60-minute idle TTL. Override it per request with the ttl request extra, or at load time with lms load <model> --ttl 3600. Models loaded explicitly with lms load have no TTL and stay until unloaded.

Auto-evict

By default LM Studio keeps at most one JIT-loaded model in memory, evicting the previous one before loading the next. Turn auto-evict off in the server settings to keep several resident at once.

The state metadata (loaded / not-loaded) shows which models are resident right now; Model Management lets you change that from Ruby.