Tuned Tensor Documentation

Tuned Tensor turns a behaviour spec into a fine-tuned small open-weight model and a paired report showing whether it beat the run baseline on representative evaluation cases.

Start with Local Training to run on a compatible NVIDIA GPU on Linux without a Tuned Tensor account. Use the Managed Quickstart when you want Tuned Tensor to operate the training infrastructure. Both paths use tunedtensor.json as the behaviour spec.

Key Concepts

Behaviour Specs

A behaviour spec is a structured description of what you want your model to do. In CLI and local workflows, this lives in your project as tunedtensor.json. It includes:

System prompt — the persona and role of the model
Guidelines — rules the model should follow
Constraints — things the model must not do
Examples — input/output pairs demonstrating desired behaviour
Base model — which open-weight model to fine-tune

Runs

A run snapshots the behaviour spec, prepares training and evaluation data, fine-tunes the model, and compares the run baseline with the tuned candidate on representative cases. Local and managed runs differ in who operates the training infrastructure.

Paired Evaluation

Tuned Tensor evaluates the baseline and tuned candidate on matched representative cases and reports score and pass-rate deltas plus per-example regressions. TT Local creates a holdout when the spec has enough examples; prebuilt data uses a test or validation split by default. Results are evidence on those cases, not a guarantee of production improvement.

Local Serving

Fine-tuned artifacts can be downloaded and served locally with the tt CLI. Use tt models serve to run a completed model with the behaviour spec prompt applied. For local-first training, evaluation, and local run tracking, use the standalone tt-local runner documented in Local Training.

Workflow

Define — Create a behaviour spec describing what your model should do
Run — Choose local or managed compute, evaluate the baseline, fine-tune the candidate, and evaluate it on matched cases
Inspect — Review the paired report, aggregate deltas, and per-example regressions
Serve — Use tt-local models serve for a local model or tt models serve for a managed artifact
Iterate — Refine your spec and run again

Managed Base Models

Managed fine-tuning currently supports a restricted set of production base models. Prices below are per 1M training tokens, per epoch. See the Billing & credits page for details on how managed runs are charged.

Model	Size	Price / 1M tokens · epoch
`google/gemma-4-E2B-it`	E2B	$0.45
`Qwen/Qwen3.5-2B`	2B	$0.45
`ibm-granite/granite-3.3-2b-instruct`	2B	$0.45
`Qwen/Qwen3-VL-2B-Instruct`	2B VL	$0.55
`meta-llama/Llama-3.2-3B-Instruct`	3B	$0.55
`bigcode/starcoder2-3b`	3B	$0.55
`google/gemma-4-E4B-it`	E4B	$0.70
`Qwen/Qwen3.5-4B`	4B	$0.70
`microsoft/Phi-4-mini-instruct`	3.8B	$0.70

CLI & API

The managed workflow is accessible through the open-source tt CLI or the REST API. For local training and evaluation, use the open-source tt-local runner.

Managed CLI — Install with npm install -g @tuned-tensor/cli. Open source (MIT) on GitHub. See the Managed CLI docs for installation and usage.
REST API — Base URL: https://tunedtensor.com/api/v1. All endpoints accept Authorization: Bearer <api-key> API key authentication or session cookies from the dashboard.