Reasoning¶

LM Studio honors OpenAI's reasoning_effort on /v1/chat/completions and answers with reasoning_content and a reasoning_tokens count, which RubyLLM reads into Message#thinking and response.tokens.thinking:

chat = RubyLLM.chat(model: 'qwen/qwen3-4b', provider: :lms)
response = chat.with_thinking(effort: :low).ask('What is 2 + 2?')

response.content              # => "4"
response.thinking.text        # => "The user is asking a simple arithmetic question..."
response.tokens.thinking      # => 40

Reasoning works on every chat protocol. On :anthropic it travels as Anthropic's thinking request field and comes back as a thinking content block; see The Anthropic API.

What a model supports¶

chat.with_thinking with no arguments, and chat.with_thinking(false), both resolve against the model registry: the provider reads each model's capabilities.reasoning from LM Studio's native listing and records it as RubyLLM reasoning options, so RubyLLM knows which efforts the model takes and whether it can be switched off at all.

model = RubyLLM.models.find('qwen/qwen3.8-27b', provider: :lms)
model.supports?(:reasoning)                  # => true
model.reasoning_option_values(:effort)       # => ["none", "low", "medium", "xhigh"]

Pass provider: :lms: without it, find searches the whole registry and can return another provider's entry for the same model name, with different reasoning options.

That information comes from the native listing, so it is only in RubyLLM.models for models present in the catalog you registered — see The Packaged Catalog.

A model that reports no way to stop reasoning — gpt-oss lists only low/medium/high — makes with_thinking(false) raise when the request is sent, on ask. That is the registry telling you the truth about the model, not a gap in the catalog.

Two vocabularies, translated for you¶

	Accepted values
`/v1/chat/completions`, `/v1/responses`, `/api/v0/chat/completions`	`none`, `minimal`, `low`, `medium`, `high`, `xhigh`
`/api/v1/chat` (`:native_chat`)	`off`, `low`, `medium`, `high`, `xhigh`, `on`
RubyLLM's registry (what `Model#reasoning_options` reports)	the first row, plus a `:toggle` option where LM Studio offers `on`

The provider translates LM Studio's off to RubyLLM's none when reading the listing, and back to off when :native_chat sends a request.

Models without reasoning¶

The surfaces differ in how they treat a model that has none. The OpenAI endpoints accept reasoning_effort for any model and silently ignore it; POST /api/v1/chat rejects the request outright:

Model 'bible-study-phi3-mini' does not expose reasoning configuration.

So on :native_chat, only ask for reasoning from a model whose listing reports it. with_thinking and with_thinking(false) refuse before any request is sent (on ask) for a model that reports no reasoning options — one in the registry without them, or one not in your catalog at all. Only an explicit with_thinking(effort:) reaches the server and gets the error above.