Your local agent. Train and run models.

Use tt to work locally or in the cloud, inspect runs, and track usage. The agent runs on your machine.

Your TT access token includes model access for the agent, with no separate model key to set up. Or bring your own provider key and choose your model.

Direct local commands work without an account.

Install
$ curl -fsSL https://tunedtensor.com/install.sh | sh

The installer places tt on your PATH. Local adapter training and serving need uv and a compatible NVIDIA GPU on Linux.

Example: inspect a run
local
tt
agent tunedtensor/managed · workflow model base
ctrl+c stop/clear · ctrl+d exit · /help commands · tab complete

Ask TT anything. Known commands run directly.
How do I inspect my latest local run?
Use runs list to find it, then runs report <run-id> to read the results.
runs list

Local agent

Inspect runs and plan your next step.

Ask TT to check your setup, explain a run, or help prepare a spec. Review proposed spec changes before applying them with /approve.

Use the same CLI for local files, cloud resources, diagnostics, and usage. The dashboard shows cloud progress and results.

Example: train and evaluate
local
init --model Qwen/Qwen3.5-2B
Created tunedtensor.json. Add your instructions and examples.
doctor
GPU, uv, and spec look ready
pipeline run --spec tunedtensor.json
pass rate  61.5% → 89.0%  (+27.5 pts)

Train and evaluate

Compare the base model with your adapter.

Define your task, examples, constraints, and base model in tunedtensor.json. TT evaluates the base and trained models on the same cases with the same scorer.

Local adapter runs set aside examples for evaluation. Prebuilt datasets require a separate validation or test split. Inspect the scores and individual regressions in the report.

Example: verify and serve
local
runs report local-8f3a
base and tuned scores · case results · evaluation gates
models verify local-8f3a
adapter checksum matches the run record
models serve local-8f3a
OpenAI-compatible  http://127.0.0.1:8000/v1

Serve your model

Verify the files and start a local server.

Local runs keep specs, adapters, events, and reports on disk. Verify the model files before serving them. To make a model active, configure a generalRegression evaluation and check that its gate passes.

Adapter serving loads the instructions saved with the run. Use --spec when serving a base model, or to supply a matching copy of an adapter's spec.

Your hardware

Supported local checkpoints.

These checkpoints support text training with LoRA adapters on CUDA. The 30B models target NVIDIA DGX Spark hardware. TT downloads weights from Hugging Face unless they are cached. Local evaluation scores the model outputs without calling a separate judge model.

Default text model

Qwen/Qwen3.5-2B

The usual starting point for classification, extraction, routing, and policy tasks.

DGX Spark

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

A 30B model for Spark hardware. Training adapts shared attention and Mamba projections; routed experts stay frozen.

Muse text training

meta-models/Muse-Glimmer-30B

Trains the text component of this checkpoint. Local examples must contain text; vision weights stay frozen.

Best fit

Tasks with clear success criteria.

Tuned Tensor is for focused application tasks. Provide representative examples, score the outputs, and decide whether the fine-tune is worth keeping.

Structured extraction into strict JSON

Classification, safety, and triage

Routing and tool selection

Format, tone, and policy adherence

Example evaluation

Email triage: test pass rate rose 27.5 percentage points.

We fine-tuned Qwen 3.5 2B on 8,000 public training examples. On a test sample of 200 examples excluded from training, pass rate improved from 61.5% to 89.0% and average score improved from 0.537 to 0.862.

These results are specific to this task and evaluation setup. Test on data that reflects your production workload.

Base pass rate

61.5%

Tuned pass rate

89.0%

Pass rate change

+27.5 pts

Test cases

200

Before you run

Check your hardware and evaluation data.

Linux and NVIDIA CUDA

Local adapter training and model serving need Linux, a compatible NVIDIA GPU, and uv. Evaluation may use CPU. Run tt hardware to check your machine.

Model support

Use a supported checkpoint from the list above. GPU memory and training settings determine which models fit your machine.

Evaluation coverage

Reports measure performance on the cases you provide. Include examples of the inputs your application will receive.

Accounts and model access

Direct local commands need no account. Use a TT token for managed agent access and cloud operations, or bring your own provider key for the agent.

Get involved

Install TT. Connect your agent.

The CLI is Apache-2.0. Sign in to create an access token, then run tt auth login to connect your local agent and cloud account. Your own provider key remains an option.