Find bad data before the trainer does.
Surface broken schemas, duplicates, leakage, and missing provenance before compute begins.
0.1.0 Beta 2 · Local-first AI model studio
Build your own AI on your Mac. Keep the proof.
Prepare the data, establish a baseline, fine-tune an open model, and measure what changed—all inside one portable project.
Tensor Cortex StudioInvoice extractionLocal project · Revision 03
Core journey · 6 of 7 / Experiment decision
Find out what actually changed.
Example answers
The same held-out example, before and after training
Invoice total · A
Expected answer
120.00
Baseline
Differs from expected102.00
After training
Matches expected120.00
Match means exact equality of the evaluated answer, not a judgment of usefulness.
This website demo uses invented answers. No model runs and no project data is read.
Compare expected, baseline, and trained answers. Export preserves the evidence; it does not certify that the model improved.
How answer review works →A training run is an event. Evidence is the product.
Open-model tools can already launch a run. The difficult part is preserving the task, dataset, baseline, method, hardware, holdout, and outcome well enough to make a defensible decision.
Keep the method with the result.
Pin the model, configuration, environment, and checkpoints so a useful run can be recovered.
Let regressions veto a pretty chart.
Compare task quality and holdout behavior; do not confuse lower training loss with a better model.
02 / One continuous project
From raw examples to a result you can defend.
Studio is designed around one ordered experiment record, not a collection of disconnected utilities.
Read the product documentation- 01Task contract
Project
Name the task, the behavior you need, and the constraints that decide whether the project is useful.
- 02Dataset revision
Data
Inspect canonical JSONL, isolate quality issues, preserve provenance, and freeze honest evaluation splits.
- 03Baseline evidence
Baseline
Measure the unchanged base model before spending time on a customization run.
- 04Recorded configuration
Recipe
Review the LoRA settings and training token budget. Apply changes deliberately, with their effect on readiness visible.
- 05Reproducible run
Train
Run a bounded MLX and LoRA experiment on Apple Silicon with the configuration recorded beside it.
- 06Decision record
Compare
Compare the same held-out examples, inspect expected and generated answers, and review changes before deciding what to use.
- 07Portable artifact
Export
Keep the adapter, evaluation summary, provenance, and usage guide together. The exact base model is required separately.
03 / The decision layer
A run can finish and still fail.
Studio lets you compare recorded results and inspect individual answers. A completed comparison permits export when its other prerequisites are satisfied, even if quality did not improve. You decide whether the adapter is useful for your task.
“Completed” describes compute. “Improved” requires evidence.
Your review checklistIllustrative decision record
QuestionEvidenceDecision
Is the data usable?Revision and health reportReady
Did the task score improve?Baseline comparisonImproved
Did anything important regress?Frozen holdoutReview
Can another machine reproduce it?Manifest and environmentRecorded
User decisionReview before useExport preserves evidence, not a quality certificate.
04 / Open project contract
The experiment leaves with you.
The canonical project lives on your Mac. Source revisions, configuration, evidence, and export intent remain inspectable instead of disappearing behind a hosted dashboard.
- Local authority
- The Mac holds the canonical project copy.
- Open artifacts
- Portable files, not a platform hostage.
- Secret-free manifests
- Credentials stay outside the project contract.
project_id: invoice-extraction
schema_version: 1
dataset:
format: jsonl
revision: sha256:41c…
holdout: frozen
base_model:
source: open-model
revision: immutable
experiment:
method: lora
backend: mlx
evidence:
baseline: required
regressions: required
provenance: includedProject portableCredentials externalSource preserved
05 / Compute without captivity
The project stays. Compute changes.
Begin with the machine already on your desk. The same project contract can later reach more compute without changing who owns it.
- 01Your MacFirst target
Apple Silicon · MLX · local project
- 02Multi-MacPlanned
Trusted local network compute direction
Local training does not require a Tensor Cortex account. Multi-Mac is a planned product direction, not a generally available capability today.
06 / Ownership boundary
Local-first is not a sync setting.
It decides which copy is authoritative.
Studio is designed so the local project remains canonical. Model downloads and update checks stay narrow, disclosed network actions rather than taking ownership of your experiment.
- Explicit data-movement boundaries
- No silent project-content transfer
- Portable model and evidence package
Latest field note
Why You Need a Baseline Before Fine-Tuning a Local LLM
A practical protocol for measuring the base model, protecting the holdout, and deciding whether a local fine-tuning experiment actually improved the task.
Read the evidence noteQuestions, answered plainly
Before you mistake direction for availability.
01What is Tensor Cortex Studio?
Tensor Cortex Studio is a local-first environment for adapting open text models on Apple Silicon. The development build connects dataset preparation, model selection, baseline evaluation, supervised fine-tuning with LoRA, example-answer comparison, and adapter export. Website download links open on 22 September 2026. Distillation and an automated method advisor are future directions.
02Is Tensor Cortex Studio available to download?
Website download links open on 22 September 2026. The signed delivery package and update feed are prepared ahead of the website launch. See the release calendar for the launch schedule.
03Is it just a graphical wrapper around training commands?
No. The development build connects model selection, dataset health, baseline evaluation, bounded training settings, example-answer comparison, and packaging in one local project. A lower training loss alone is not treated as proof that a model improved.
04Which computers will it support first?
The first product target is Apple Silicon Mac hardware and text decoder language models through an MLX-based workflow. The development build has a searchable model list and accepts compatible public Hugging Face or local MLX models outside it. Technical tests passed for 17 pinned catalog entries; minimum-memory, task-quality, and public release qualification are separate. See the support matrix for measured boundaries.
05Does my data have to leave my Mac?
Local projects keep their canonical copy, datasets, model files, checkpoints, and experiment history on the Mac. Manual catalog refresh, model metadata inspection, confirmed model downloads, and optional software-update checks are disclosed separately and do not include project workload content. Catalog refresh is explicit-only and scoped to the Studio version and build; it is not telemetry.
06Does Studio require a Tensor Cortex account?
The local Studio workflow is designed to run without a Tensor Cortex account. The project on the Mac remains the canonical copy.
07Can it train a foundation model from scratch?
No. The first beta uses supervised fine-tuning with LoRA to adapt existing text models. Model-native response formats and model-derived token budgets do not change that training method. Retrieval, preference optimization, and distillation are later product directions, not available training modes in this beta.
08What will an experiment produce?
The development build exports a LoRA adapter with configuration, evaluation summary, source identities, provenance, and USAGE.md instructions. The base model, dataset, Python runtime, and Studio application are not included. Use the exact recorded base model with a compatible MLX-LM runtime. Export does not certify task quality; review the comparison before deciding how to use the adapter.
0.1.0 Beta 2Evidence before promises
Watch Studio become real.
Tensor Cortex Studio is under active development. Website download links open on 22 September 2026; published capabilities and benchmarks are tied to reproducible product evidence. See the release calendar at https://tensorcortex.com/releases/.