The OpenAI API¶

The surface most code already targets, and the one with the fewest surprises: on Chat Completions, everything RubyLLM can express works.

Chat Completions — `:chat_completions`¶

The default protocol; nothing to configure. LM Studio honors the usual OpenAI vocabulary on POST /v1/chat/completions: model, messages, temperature, top_p, max_tokens, stream, stop, presence_penalty, frequency_penalty, logit_bias, seed, tools / tool_choice, and response_format with a json_schema (strict structured output). Vision models accept image content parts.

You reach all of it through RubyLLM's normal API — with_temperature, with_max_output_tokens, with_tools, with_schema, chat.ask(with: 'image.png') — nothing LM Studio-specific required.

Structured output is grammar-constrained

LM Studio builds a grammar from your schema, so the JSON is always structurally valid — even from weights with no tool training. Whether the values make sense is up to the model: gpt-oss models mangle json_schema values on LM Studio, so the specs use a qwen3 model for structured output.

Responses — `:responses`¶

LM Studio also serves OpenAI's newer Responses API at POST /v1/responses:

RubyLLM.configure { |config| config.lms_protocol = :responses }

Streaming, multi-turn conversations and client-side tools all work.

with_schema does not work here

LM Studio ignores the schema on this endpoint and nothing raises, so you get unconstrained prose where you expected JSON. Use :chat_completions for structured output.

with_thinking works, reasoning text included. That takes this gem's own Responses subclass: OpenAI never returns raw reasoning, only an optional summary, so RubyLLM's stock protocol reads a reasoning item's summary. LM Studio runs the model locally and has nothing to hide, so it returns the reasoning itself in content as reasoning_text (streamed as response.reasoning_text.delta). The provider reads both shapes.

Embeddings¶

Embeddings go to POST /v1/embeddings regardless of which chat protocol is in effect — the provider routes them there itself, because that is the only place LM Studio serves them:

RubyLLM.embed('Hello world', model: 'text-embedding-nomic-embed-text-v1.5', provider: :lms)

LM Studio request extras¶

LM Studio accepts request fields that are not part of OpenAI's vocabulary. RubyLLM merges with_provider_options into the payload as-is, so these work on every chat protocol except :native_chat, whose server rejects unknown keys:

Option	What it does
`ttl`	Idle time-to-live in seconds for the model handling this request. A JIT-loaded model unloads itself after sitting idle that long (default 60 minutes; each request resets the timer).
`draft_model`	Speculative decoding: a small draft model proposes tokens the main model verifies, often speeding generation up substantially. Both models must share a tokenizer family.
`top_k`	Sample only from the k most likely tokens.
`min_p`	Discard tokens below this base probability.
`repeat_penalty`	Penalize recently generated tokens to discourage repetition.

chat = RubyLLM.chat(model: 'qwen/qwen3-4b', provider: :lms)
chat.with_provider_options(
  ttl: 300,                        # unload after 5 idle minutes
  draft_model: 'qwen/qwen3-0.6b',  # speculative decoding
  top_k: 40,
  min_p: 0.05,
  repeat_penalty: 1.1
)
chat.ask('Hello')