Tinker gives you a tuning API. Tuned Tensor gives you the whole closed loop — behaviour spec, fine-tune, evaluate, auto-tune — as an MIT-licensed CLI you run and inspect yourself.
How it compares
Tinker hands you low-level training primitives and full control — you write the loop, the evals, and the iteration. Tuned Tensor is the opinionated layer on top: describe behaviour in a spec, and it compiles, augments, fine-tunes, evaluates, and auto-tunes for you, then hands you a small local model.
Tuned Tensor | Raw APIRaw tuning API (e.g. Tinker) | |
|---|---|---|
| Describe behaviour in a structured spec | Write a training loop | |
| No training-loop or GPU code to write | — | |
| AI-augments your examples into a training set | — | |
| Built-in evaluation every run (LLM-as-judge) | DIY (cookbook) | |
| Regression detection across runs | DIY | |
| Auto-tune loop from evaluation feedback | — | |
| One-command export to GGUF / Ollama | Download weights | |
| Agent-readable workflow (/skill.md) | — | |
| Full low-level control at research scale (1B–1T+) | Small models (2–4B) | |
| Managed training infrastructure |
Both are managed services with open-source pieces: the Tuned Tensor CLI is MIT-licensed, and Tinker ships an open-source cookbook. The difference is the workflow — a behaviour spec and a built-in closed loop versus a low-level training API you build on.
Tuned Tensor turns a behaviour spec into a tuned open-weight model, then uses regression reports to improve the next run.
01
Rules, constraints, examples, and target base model.
02
Compile training data and tune a small open-weight model.
03
Score outputs, inspect failures, and compare against baselines.
04
Use AI feedback to improve the spec and start the next run.
Auto-tune can repeat the loop until the target score is reached or the iteration limit is hit.
Use case story
We fine-tuned Qwen 3.5 2B on a public 10k-row dataset to triage email into a strict JSON object — triage action, priority, and prompt-injection risk — then measured the gains against the base model on held-out test examples.
Test pass rate
61.5%89.0%
Test avg score
0.5370.862
Dataset
—10,000 rows
Explore the documentation to see how behaviour specs, runs, and evaluations work under the hood.