Fine-tune an open model. See if it got better.
tt is one local terminal: a laptop agent, CUDA fine-tuning on your GPU, and a paired report that compares the tuned model with its baseline. No account.
$ curl -fsSL https://tunedtensor.com/install.sh | shThe installer places tt on your PATH. Training still needs uv and a compatible NVIDIA GPU on Linux. Read the CLI docs.
tt v0.13.0
agent anthropic/claude-sonnet-4-5 · workflow model base
ctrl+c stop/clear · ctrl+d exit · /help commands · tab complete
Ask TT anything. Known commands run directly.
› What happened in my latest training run?
Pass rate rose 27.5 pts on the held-out set. Two regressions in billing replies.
› runs list▊Local agent
Ask TT about the run, not the training loop.
The harness and conversation stay on your laptop. Inference uses the provider you pick. /model switches the agent model; it does not change the fine-tune base.
Tools are Tuned Tensor only: specs, runs, diagnostics, and reports. Spec writes wait for /approve. There is no cloud token and no general filesystem access.
› init --model Qwen/Qwen3.5-2B
✓ Created tunedtensor.json — add your examples
› doctor
✓ GPU, uv, and spec look ready
› run
pass rate 61.5% → 89.0% (+27.5 pts)Train and evaluate
Every run is a baseline-versus-tuned test.
tunedtensor.json is the behaviour spec: task, examples, constraints, and base model. The same cases and scorer run against the baseline and the candidate.
The local workflow creates a holdout when the spec has enough examples. For prebuilt data it requires a test or validation split by default. Loss is not the product; the paired report is.
› runs report local-8f3a
2 regressions · 11 improvements · gate passed
› models verify local-8f3a
✓ adapter checksum matches the run record
› models serve local-8f3a
OpenAI-compatible http://127.0.0.1:8000/v1Keep or iterate
Verify the artifact. Serve it on hardware you control.
Specs, adapters, events, and JSON reports stay on disk. Activation is optional and waits on a configured generalRegression gate.
tt models serve exposes an OpenAI-compatible endpoint. Pass --spec when the server should enforce the same instructions used in training.
Your hardware
Certified local models. Apache-2.0 CLI.
Text SFT with LoRA/PEFT on CUDA. The 30B paths are intended for NVIDIA DGX Spark-class unified memory. Downloads use Hugging Face unless the weights are already cached. OpenRouter is optional for LLM-judge scoring.
Default text SFT
Qwen/Qwen3.5-2B
The usual starting point for classification, extraction, routing, and policy tasks.
DGX Spark 30B
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
MoE path for Spark-class unified memory. Adapts shared attention/Mamba projections, not every routed expert.
Vision-language
meta-models/Muse-Glimmer-30B
Fine-tunes the text tower; the vision tower stays frozen. Use when the task is image-text to text.
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 decisions
Format, tone, and policy adherence
Measured on held-out data
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 200-example held-out test sample, 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
Held-out cases
200
Constraints
What tt is not.
Linux and NVIDIA CUDA
Training needs a compatible NVIDIA GPU and uv. Evaluation and serving may use CPU. Apple silicon training is not certified.
Small open models
Tuned Tensor is for focused application tasks, not frontier-scale research. Certified checkpoints are listed below.
Evidence, not a production guarantee
Reports summarise the cases you provide. They do not promise live-traffic performance.
Hosted training is retired
The CLI is local-only. No Tuned Tensor account, API key, or cloud run is required.
Get involved
Install tt. Read the source.
The CLI is Apache-2.0. Training, evaluation, and serving run on hardware you control.