In development · Local-first AI model studio

Your Mac is an AI lab. Make it usable.

Prepare datasets, fine-tune and evaluate open models, compare real results, and package what works—locally first, with cloud compute only when you choose it.

Follow the build Explore the workflow

Local-first and private by default
Baseline before training
Free on your Mac
Portable models and experiments

Experiment workspace product direction

Apple Silicon MLX local project

Inputs

dataset 18,420 labeled examples JSONL · local

base model Open 4B model pinned revision · MLX

target Structured extraction schema-valid output · measurable

Experiment

Measure dataset health baseline eval QLoRA · holdout

Evidence

baseline task score · recorded recorded

candidate candidate · compared review

package LoRA + experiment card portable

preflight → dataset revision 3 duplicates isolated · source preserved

training → local M-series Mac checkpointed · budget $0

holdout → regression found package held until the evidence is clean

The workflow gap

Open-model tooling is powerful. The workflow is fragmented.

Building a useful custom model still means stitching together dataset scripts, trainer configuration, hardware estimates, evaluation notebooks, checkpoints, and export tools. Tensor Cortex Studio turns them into one evidence-led project.

Data fails quietly

Duplicates, broken schemas, contamination, and missing provenance can invalidate an experiment before training begins.

The method is guessed

Fine-tuning is often started when retrieval, a better prompt, or a smaller task-specific adapter would be the better choice.

Loss looks like success

A falling training curve says little about task quality, regressions, target-device latency, or whether the result is portable.

01 / Studio workflow

One project from raw data to a portable model.

The same manifest carries the task, dataset revisions, base model, training configuration, eval suites, hardware measurements, and output package. The project stays coherent whether compute runs on one Mac or later moves to cloud.

01DefineTask, target, constraints
02PrepareDataset health and splits
03BaselineMeasure before training
04TrainLocal MLX experiment
05CompareHoldout and regressions
06PackageModel, card, provenance

Data Studio

Know your data before you spend a training hour.

Import the formats technical teams already use, inspect quality and provenance, then create an explicit train, validation, holdout, and stress split. Every fix creates a new revision; source data is never silently rewritten.

Schema and quality checksMissing outputs, duplicates, length, balance, and format.
Leakage protectionTrain/eval overlap and holdout contamination become visible.
Provenance by revisionEvery dataset change remains inspectable and reproducible.

Dataset health · 18,420 examples

LocalSource preserved · revisioned fixes

311 duplicates isolated. 42 missing outputs need review. Eight train/eval overlaps block the baseline.

Create clean revision Review 361 issues

Training Studio

Train locally without turning your Mac into an infrastructure project.

Start with a measured baseline and a hardware-aware plan. Guided mode proposes a bounded configuration; expert mode keeps the important MLX and LoRA controls visible, versioned, and explainable.

MLX-first executionDesigned first for text models on Apple Silicon.
Hardware-aware preflightMemory, runtime, and risk estimates before the run.
Checkpoint and resumeExperiment state survives interruption and stays reproducible.

Read the development documentation

ExperimentPreflightEvidence

project: invoice-extraction
task: structured-output
base_model:
  family: open-4b
  revision: pinned
method:
  trainer: mlx
  strategy: qlora
evaluation:
  baseline: required
  holdout: protected
compute:
  target: this-mac

DatarevisionedBudgetlocal · $0

Evaluation & packaging

Prove the model improved before you ship it.

Compare the baseline, candidate, and reference model on the same versioned eval. Task success, format validity, regressions, memory, and latency travel with the result so a checkpoint never becomes an evidence-free artifact.

Task-specific evaluationSchema, exact-match, classification, tool, and custom checks.
Protected holdoutRegression gates stay outside training and tuning decisions.
Portable packageModel or adapter, experiment card, provenance, and known limits.

COMPARE

Baselinebase model · recorded

CandidateQLoRA adapter · revision 7

REQUIRE

Holdoutno blocked regression

Outputvalid JSON schema

EVIDENCE

Baseline: recorded
Candidate: improved
Holdout: review required
Package: held

Illustrative product direction · not a benchmark claim

02 / Local-first boundary

Your project belongs on your Mac.

The Mac keeps the canonical project, dataset revisions, experiment history, and artifacts. Local work requires no cloud account. A future cloud job is an explicit, budgeted compute choice—not a silent migration of project ownership.

Review the security architecture

Project authorityThis Mac

Local experimentdataset and model stay on device

default

Managed cloud computeregion, estimate, and hard budget shown first

optional

Portable by design No project lock-in Open manifests and model formats keep datasets, recipes, evals, and artifacts usable outside Tensor Cortex Studio.

One model-building environment

The parts of a serious experiment, connected.

AdvisorPrompt, retrieval, fine-tuning, preference, or distillation

DataQuality, splits, provenance, revisions, and contamination checks

ModelsLicenses, revisions, compatibility, memory, and target device

TrainingGuided and expert MLX experiment configuration

EvaluationBaseline, holdout, regression, variance, and task metrics

RecoveryCheckpoints, resumable jobs, and explicit failure state

EvidenceExperiment cards, hardware measurements, and known limits

PackagingMLX, LoRA, safetensors, GGUF, and Ollama directions

03 / Compute continuity

Start on your Mac. Add compute only when the experiment needs it.

Tensor Cortex Studio is under active development. There is no public app download or generally available cloud service yet; published capabilities and benchmarks will be tied to reproducible product evidence. Local execution is the first product boundary; Multi-Mac, managed cloud, and BYOC remain later stages and are shown here as direction.

Canonical project Dataset · Recipe · Eval Stored on your Mac

Experiment environment Tensor Cortex Studio Plan · Train · Compare · Package

DefaultThis Mac · free

When chosenMulti-Mac · cloud · BYOC

This Mac · First targetApple Silicon, text models, MLX, SFT and LoRA; public builds are not available yet.

Multi-Mac · PlannedTrusted local workers will extend the same project without creating a separate platform.

Tensor Cortex Cloud · PlannedManaged jobs will show region, time and cost estimates, checkpoints, and a hard budget before submission.

BYOC · FutureHeavy users may connect supported provider accounts while keeping the Tensor Cortex project contract.

Evidencebeforetraining

Define the baseline and success criteria before spending compute.

Localbeforecloud

Use the Mac first; move only when time or memory makes it worthwhile.

Projectbeforeframework

Experiments remain coherent even when trainers and compute targets change.

Portabilitybeforelock-in

Your data, recipes, evals, models, and evidence remain exportable.

04 / FAQ

Tensor Cortex Studio questions, answered.

The practical product, privacy, availability, and model-building boundaries.

What is Tensor Cortex Studio?

Tensor Cortex Studio is a local-first environment for building task-specific open models on Apple Silicon. It is designed to help technical users prepare data, choose an appropriate customization method, establish a baseline, fine-tune or distill a model, evaluate the result, and export a portable package.

Is Tensor Cortex Studio available to download?

Not yet. The product is under active development and there is no public application build today. The website documents the intended product boundary and will publish download, compatibility, and benchmark information only when supported by tested release evidence.

Is it just a graphical wrapper around training commands?

No. The product is intended to connect method selection, dataset health, baseline evaluation, hardware-aware planning, training, regression checks, and packaging in one reproducible project. A lower training loss alone is not treated as proof that a model improved.

Which computers will it support first?

The first product target is Apple Silicon Mac hardware and text decoder language models through an MLX-based workflow. A deliberately small supported model and task matrix will expand only after compatibility is measured.

Does my data have to leave my Mac?

Local projects are designed to keep their canonical copy on the Mac, and local training does not require a cloud account. If a user later chooses managed cloud compute, the interface will show the region, estimated cost, data movement, and hard budget boundary before a job is submitted.

Will local use be free?

The planned core Studio application, local training, and Multi-Mac compute are free. The commercial model is based on optional managed cloud compute and paid orchestration for users who connect their own cloud accounts. Final pricing will be published before those services become available.

Can it train a foundation model from scratch?

That is not the initial product. Tensor Cortex Studio focuses on adapting existing open models with methods such as prompt design, retrieval, supervised fine-tuning, preference optimization, and distillation, then measuring whether the adapted model is better for a defined task.

What will an experiment produce?

The intended output is more than a checkpoint: a portable model or adapter package accompanied by its dataset and model revisions, experiment configuration, evaluation comparison, hardware and cost measurements, provenance, and known limitations.

Tensor Cortex Insights

Evidence-led field notes for local model builders.

Practical guides on datasets, evaluation, Apple Silicon, open-model fine-tuning, distillation, packaging, and the decisions that make an experiment trustworthy.

View all Insights

Dataset and evaluation Updated August 29, 2026

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 article

05 / Build updates

Follow Tensor Cortex Studio as it becomes real.

Tell us what you want to build and which Apple Silicon Mac you use. We will share material development milestones, compatibility evidence, and early-access openings.

Real progressMilestones and release evidence, not a manufactured launch countdown.
Measured compatibilitySupported Macs, models, tasks, and limits published as they are verified.
Early accessInvitations only when a build can support a useful end-to-end workflow.

Prefer email? hello@tensorcortex.com