ruby_llm-providers-lms¶

A RubyLLM provider for LM Studio: run local models through the same Ruby API you use for hosted ones.

gem 'ruby_llm-providers-lms', require: 'ruby_llm/providers/lms'

chat = RubyLLM.chat(model: 'qwen/qwen3-4b', provider: :lms)
chat.ask('Hello').content

No API key and no configuration: start LM Studio's server (lms server start) and the :lms provider talks to http://localhost:1234/v1.

Get started Judgments GitHub

What it covers¶

LM Studio serves several APIs on one port, and this gem speaks all of them:

API	What you get	Protocols
OpenAI	Chat, streaming, tools, structured output, vision, reasoning, embeddings	`:chat_completions` (default), `:responses`
Anthropic	Claude-style Messages API against a local model	`:anthropic`
Native REST	Inference stats, server-stored conversations, MCP integrations, model management	`:native_v0`, `:native_chat`
Decisions	Typed judgments from decision models: probabilities, choices and scores in milliseconds	`:decisions` (default), `:system_one`

How requests reach LM Studio

New in 0.3: judgments with decision models¶

LM Studio 0.4.26 added endpoints for decision models: models that answer typed questions with probabilities instead of writing text. RubyLLM 2.1 calls these judgments, and this gem points them at LM Studio:

class TicketTriage < RubyLLM::Judge
  model 'd1-omni-600m', provider: :lms

  probability :urgent, 'Does this need attention today?'
  choice :team, 'Which team should handle this?' do
    billing   'Charges and refunds'
    technical 'Bugs and outages'
  end
  score :frustration, 'How frustrated is the customer?', %w[Calm Annoyed Angry]
end

judgment = TicketTriage.judge('You charged my card twice. Refund it today.')
judgment.urgent.probability  # => 0.97
judgment.team.choice         # => :billing

Open-source decision models such as D1 Omni 600M answer in tens of milliseconds on a laptop. See Judgments.

The one thing to know first¶

What a call does depends on two choices, not one

LM Studio runs whatever model you downloaded behind every protocol. The protocol decides what is expressible — with_schema is honored on :chat_completions and silently ignored on :responses. The model decides whether it honors what the protocol accepted — qwen3 obeys tool_choice: "required", gpt-oss ignores it. Test against the models you intend to ship on. Choosing a protocol maps the first axis.

Requirements¶

Ruby 3.3 or newer
RubyLLM 2.1.1 or newer
LM Studio with its server running (0.4.26 or newer for judgments)