Tuned Tensor

Tuned Tensor documentation

Use the tt CLI to work with a local agent, train and evaluate models, inspect runs, and check usage. Use tt pipeline run to train on your hardware or an existing GPU instance in your AWS account. See Run setup. Use tt cloud for account resources, labeling, and historical runs.

Direct local commands need no Tuned Tensor account or token. For the managed agent, save a TT access token with tt auth login. The same token authorizes cloud commands. The web dashboard shows cloud progress, published reports, and account settings.

You can also use your own OpenRouter key and choose any model available to that account. Tuned Tensor selects the model for managed inference; your own key gives you control of the model. See Agent for setup.

Core concepts

Behaviour specs

A behaviour spec is a structured description of what you want your model to do. It lives in your project as tunedtensor.json. It includes:

  • system_prompt: the model's role and instructions.
  • guidelines: rules the model should follow.
  • constraints: behaviour the model should avoid.
  • examples: inputs and expected outputs.
  • base_model: the model to fine-tune.

Runs

An adapter run saves a copy of the spec, prepares training and evaluation data, trains a LoRA adapter, and compares the result with the base model. Run reports stay local, including when GPU work runs on AWS. Use tt publish <local-run-id> if you want to upload run evidence to Tuned Tensor and display it in the dashboard.

Paired evaluation

TT evaluates the base and tuned models on the same cases and reports changes in score and pass rate, with results for each example. Local adapter runs set aside some inline examples for evaluation. With a prebuilt dataset, TT uses the validation split, or the test split if no validation split is provided. Choose evaluation data that represents the work you expect the model to handle.

Local serving

Use tt models serve <model-id> to serve a verified model through an OpenAI-compatible API. To use tt serve active, first activate a model from a run that passed the configured generalRegression checks. See Local training for setup and hardware requirements.

Workflow

  1. Create a behaviour spec with representative inputs and expected outputs.
  2. Run the pipeline to train the model and compare it with the baseline.
  3. Review the report, including examples where the tuned model performed worse.
  4. Serve the model with tt models serve <model-id>.
  5. Update the spec or data and run again as needed.

Supported base models

The local adapter runtime supports text SFT with LoRA/PEFT on CUDA:

  • Qwen/Qwen3.5-2B: the default for most application tasks.
  • nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16: intended for DGX Spark or hardware with comparable unified memory.
  • meta-models/Muse-Glimmer-30B: trains the text tower while keeping the vision tower frozen.

See Model recommendations for when to pick each one.

CLI

Install with curl -fsSL https://tunedtensor.com/install.sh | sh, or npm install -g --ignore-scripts @tuned-tensor/cli. Open source (Apache-2.0) on GitHub. See the TT CLI docs and the Quickstart.