Ollama¶

from datafast import ollama

model = ollama()                                  # gemma4:12b, on your own machine
small = ollama("qwen3:0.6b", thinking=False)

Every field on Served models works here. This page covers what is specific to Ollama.

Supported models¶

The served-model catalog lists no Ollama model. On a hosted provider the model name tells datafast what it is talking to. Here it does not: the model is whatever you pulled onto the machine, and the name is only a label you chose when you pulled it. So every Ollama model resolves by what its name contains:

Each name is a whole model family that thinks, never a family where only some models do. A wrong guess here sends a reasoning parameter to a model that rejects it, so a family gets added only when all of it thinks. A single thinking model whose name says nothing — llama3.2 say — gets the plain chat profile, and the probe below is how you find out.

Three models the rest of this page uses:

Note gemma3 and gemma4: one digit apart, different profiles, because gemma4 thinks and gemma3 does not.

The default is a big model. gemma4:12b is around 7.6 GB and has to fit on your machine, which is not true of any other provider's default. On a smaller machine pass a lighter id — gemma3:4b at ~3.3 GB, or qwen3:0.6b at ~0.5 GB if you want reasoning.

The two profiles¶

thinking=True and thinking=False work on the reasoning profile too; they are the portable controls, and each resolves to one of the values below.

Image is declared for every Ollama model, and most cannot see. The profiles have no way to know which weights you pulled, so they declare image support for all of them. Datafast will happily send an image to a text-only model and the daemon will reject it. Vision models such as gemma3, gemma4 and llama3.2-vision do work. If you are not sure, probe.

Ask the daemon: probe_capabilities()¶

Because the pulled model is a property of the machine, the profiles are a guess. The daemon knows the exact answer, so probe_capabilities() asks it:

from datafast import ollama

model = ollama("qwen3:0.6b")
# model.probe_capabilities() -> frozenset({"completion", "thinking", "tools"})

It returns Ollama's own capability names — completion, vision, audio, thinking, tools, embedding, insert — not datafast's vocabulary. Use it to check the two things the name cannot tell you: whether the model can really see, and whether it can really think.

It is opt-in and nothing calls it for you. It makes an HTTP call to the daemon, at the same address your generate calls use, and raises if the model is not pulled or no daemon is listening.

Reasoning¶

thinking=True sends reasoning_effort="low". gpt-oss honours the level, so low, medium and high mean something there; every other thinking model treats any level as plain on.

thinking=False sends think=False, which is Ollama's own parameter rather than an effort. Two reasons it is done that way. Leaving the parameter out would not turn reasoning off — qwen3 thinks by default. And LiteLLM's usual route, mapping the effort to a string, sends that literal string for gpt-oss, which Ollama rejects. Passing think directly avoids both.

frequency_penalty is not offered¶

Ollama's repetition control is repeat_penalty: a multiplier that is neutral at 1.0, where values below 1.0 reward repetition. LiteLLM renames frequency_penalty onto it without rescaling the number, so an OpenAI-style 0.15 would arrive as strong encouragement to repeat. Datafast therefore leaves frequency_penalty out of both profiles. Pass repeat_penalty on its own scale instead — it goes through as a provider parameter:

from datafast import ollama

model = ollama("gemma4:12b", top_p=0.85, repeat_penalty=1.2)

This is an Ollama quirk. A self-hosted vLLM or llama.cpp server reached through openai_compatible keeps the usual meaning of frequency_penalty.

Batching¶

Ollama has no batch endpoint, so datafast issues concurrent single calls and warns that it is doing so. Expected here, not a misconfiguration. Tune it with max_concurrent (default 4). One local daemon serves one model at a time, so raising it far above the default buys little.

Where to go next¶

• Served models — every configuration field, shared by all providers.
• Model defaults — the default model id of every provider.
• Installation — OLLAMA_API_BASE and the .env rules.
• Served models guide — narrative introduction.
• API reference — generated signatures.