Not possible to pass `think = FALSE` to Ollama models via `chat_ollama()`

[AdaemmerP]

I’m using chat_ollama() with a local Ollama model (qwen3.5:4b). Ollama allows to disable reasoning for thinking models using the think = false parameter (CLI flag --think=false). The provided Python code works. However, I haven’t found a way to pass this parameter through chat_ollama() with ellmer.

From looking at the implementation, it seems that:

• chat_ollama() uses the OpenAI-compatible API (/v1/chat/completions).
• The chat_params() method for ProviderOllama maps only a limited set of parameters (temperature, top_p, max_tokens, etc.):

method(chat_params, ProviderOllama) <- function(provider, params) {
  standardise_params(
    params,
    c(
      frequency_penalty = "frequency_penalty",
      presence_penalty = "presence_penalty",
      seed = "seed",
      stop = "stop_sequences",
      temperature = "temperature",
      top_p = "top_p",
      top_k = "top_k",
      max_tokens = "max_tokens"
    )
  )
}

I added think = "think" locally to the package and re-installed it, but that did not change anything. I also tried passing the argument via api_args:

chat <- chat_ollama(
  model = "qwen3.5:4b",
  api_args = list(think = FALSE)
)

but this also does not seem to propagate the parameter to the Ollama request.

Is there currently a supported way to pass think = FALSE to Ollama models when using chat_ollama() or is this limitation due to using the OpenAI-compatible endpoint instead of the native Ollama API?

[zilinskyjan]

Unfortunately I don't have a solution, but wanted to report that:

• I've had similar issues with both chat_ollama() and chat_openrouter()
• Sometimes I was able to turn off the reasoning trace: so it'd be excluded from the output but a model like Qwen 3.5 would still be reasoning internally...

Hoping to look into this more later, but would be curious if you learn something, @AdaemmerP.

[bastianolea]

I'm having the same issue. Running in Ollama:

% ollama run qwen3:14b --think=false
>>> what do you think of R, in a few words
R is powerful, flexible, and widely used in data science and statistics.

Answers are instant.

But trying the same with {ellmer} yields thinking (and thus a long delay before answers) every time:

chat <- chat_ollama(model = "qwen3:14b",
                    api_args = list(think = FALSE), # also tried with "false"
                    params = params(reasoning_tokens = 0,
                                    reasoning_effort = "low")
)

chat$chat("what do you think of R, in a few words")

previous <- chat$last_turn()

str(previous)

previous@contents

reasoning <- previous@json$choices[[1]]$delta$reasoning
nchar(reasoning)

Over a 1600 characters of reasoning. Changing reasoning_effort changes almost no difference in reasoning length. Anybody knows how to disable thinking?