Where it sits

There are good tools on every side of this one. The honest comparison is about where your data goes, where the pipeline starts and where it stops.

Hosted fine-tuning

AutoTrain, Together, Fireworks and the OpenAI fine-tuning API are excellent when the corpus can travel. When it cannot, they are not an option. SLM never uploads a file; the trade is that you bring the GPU.

Model runners

Ollama and LM Studio run models beautifully. They do not make them. SLM is the step before: the corpus becomes a dataset, the dataset becomes a bundle, and the bundle is what you would then run.

Training recipes

Unsloth, Axolotl and LLaMA-Factory are fast, mature fine-tuning stacks for people who already have a clean dataset and their own tool-calling harness. SLM starts one step earlier with curation, exclusions and review, and ends one step later with the tool loop, its schema validation and its confirmation tiers.

What SLM is not

It is not a frontier model and does not pretend to be. The scratch lane is about a million parameters and answers from retrieval. The fine-tune lane, when it lands, adapts a small open instruct model, not a large one. If you need a chatbot about the world, use one. If you need a model about your world that can touch your tools under your rules, this is that.