Skip to documentation

Documentation

Build a model project around evidence, not trainer folklore.

The intended Tensor Cortex Studio workflow for choosing a method, preparing data, establishing a baseline, running a local experiment, evaluating the candidate, and packaging a portable result.

Status: product under active development; no public app build

Reviewed:

Current availability boundary

Tensor Cortex Studio is not available for public download today. These pages document the intended product contract and first-release boundary. A capability becomes a supported claim only after an executable artifact, compatibility test, or reproducible benchmark establishes it.

Start by choosing the right intervention

Fine-tuning is one option, not the default answer to every model problem. The Studio advisor is intended to distinguish changing knowledge, changing behavior, task specialization, output format, preference, and compression before it recommends compute.

When to consider each model customization method
NeedLikely starting pointFirst evidence
Frequently changing knowledgeRetrieval and a behavior evalAnswer quality and citation behavior
Stable task or output schemaSFT or LoRATask baseline and protected holdout
Style or preferencePrompt, SFT, or preference optimizationPairwise or rubric-based comparison
Smaller task-specific modelResponse distillationTeacher, baseline, and student comparison

The Studio workflow

  1. Define the project. Record the task, target device, privacy boundary, latency, memory, budget, and measurable success criteria.
  2. Prepare the data. Validate schema, duplicates, length, PII and secret risk, class balance, provenance, license, and split integrity.
  3. Establish the baseline. Run the exact evaluation that will later judge the candidate; protect the holdout from training and tuning.
  4. Plan the experiment. Pin the base model and revision, choose the method, estimate local memory and time, then review the configuration.
  5. Train and recover. Run on the Mac first, preserve checkpoints, and keep interruption state explicit.
  6. Compare and package. Review task quality, regressions, latency, memory, cost, provenance, and known limits before exporting.

Deliberately narrow first-release scope

  • macOS on Apple Silicon.
  • Text decoder language models through an MLX-based workflow.
  • Dataset import from JSONL, CSV, Parquet, and compatible Hugging Face sources.
  • Dataset health, train/validation/holdout separation, and baseline evaluation.
  • Supervised fine-tuning and LoRA, with guided and expert configuration surfaces.
  • Checkpoint and resume, candidate comparison, and MLX or LoRA adapter export.
Outside the initial release: general chat, agent orchestration, image/audio/video, foundation-model pretraining, arbitrary cloud code, enterprise SSO, and a broad marketplace.

Documentation map

Product

Studio overview

See the product workflow, local-first boundary, compute direction, and availability state.

Explore the workflow →

Data

Local data and security

Understand project ownership, local processing, optional cloud transfer, and website data handling.

Review data boundaries →

Machines

Project and compatibility contract

Review the intended project manifest and future version/compatibility API boundary.

Read the compatibility reference →

Research

Tensor Cortex Insights

Read evidence-led field notes on local models, datasets, evaluation, training, and packaging.

Browse Insights →

Help shape the supported first-release matrix.

Tell us your Mac, model size, task, dataset shape, and biggest workflow blocker.

Follow the build