Current availability boundary
The public beta is available to download from the website as of 22 September 2026. See the release calendar for planned and actual publication dates.
Tensor Cortex Studio build 8 is available as a public beta. 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
The first beta implements supervised fine-tuning with LoRA. An automated method advisor, retrieval, preference optimization and distillation remain future directions. The table below is general method-selection guidance, not a list of training modes available in Studio.
| Need | Likely starting point | First evidence |
|---|---|---|
| Frequently changing knowledge | Retrieval and a behavior eval | Answer quality and citation behavior |
| Stable task or output schema | SFT or LoRA | Task baseline and protected holdout |
| Style or preference | Prompt, SFT, or preference optimization | Pairwise or rubric-based comparison |
| Smaller task-specific model | Response distillation | Teacher, baseline, and student comparison |
The Studio workflow
Follow the practical seven-step guide for dataset examples, model selection, settings, progress, answer comparison, adapter use and project history.
- Project. Create or open the local experiment workspace.
- Data. Prepare train, valid and test JSONL files. Preview records, then run Inspect Data for full inspection and a recorded revision.
- Baseline. Select and compatibility-test the model, then record its answers on the held-out test split.
- Recipe. Review bounded LoRA settings and the training token budget. Apply changes with their readiness impact visible.
- Train. Follow local progress and use cancellation or recovery when offered by Studio.
- Compare. Review the same examples' expected, baseline and trained answers. Exact match alone does not establish usefulness.
- Export. Preserve the adapter, evaluation summary, provenance and usage guide. Export eligibility does not certify quality improvement.
Deliberately narrow first-release scope
The development build includes a searchable offline model list, manual public Hugging Face or local MLX model selection, and an explicit-only signed catalog refresh for each Studio version and build. Models outside the list can be inspected and compatibility-tested; the list is not an allowlist. See model selection andthe 17-model technical test boundary.
Model-native messages can retain reasoning and tool-call fields supported by the supplied template, without executing tools. Training, input, and response token budgets follow the selected model's declared context; oversized records are not silently truncated. Read thedata requirements andbudget controls and memory limits before preparing a run. These features do not imply public release qualification.
- macOS on Apple Silicon.
- Text decoder language models through an MLX-based workflow.
- Canonical JSONL dataset import.
- 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 LoRA adapter export with experiment, evaluation, and provenance evidence.
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, network requests, 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 →Support
Compatibility and diagnostics
Check the measured Mac, model, data, resource, export, and privacy boundary before testing a build.
Review the support matrix →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.