A small measured matrix is more useful than a long compatibility promise.

Current candidate and changes

0.1.0 Beta 2 build 8 was accepted by Apple notarization on 14 September 2026. The retained ZIP extraction passed strict code-signature, stapled-ticket and Gatekeeper checks. The signed delivery package and update feed are available; the public beta is available to download from the website as of 22 September 2026.

Beta 2 preserves saved model locations and evaluation references during temporary verification failures. A missing model location can be recovered from My Models only after its exact identity is verified, without a new download or silent model switch. Later candidate builds improve model-picker keyboard focus and accessible labels. Build 8 removes unused PyArrow from the distributed runtime and includes the corrected third-party notices.

These changes do not expand the supported platforms, model rights or memory guarantees. Distribution approval and final publication checks remain open.

The local text-training workflow candidate

For instructions rather than compatibility requirements, follow theseven-step local workflow guide, including dataset examples, per-example comparison, adapter usage and training history.

The model library includes a searchable offline catalog and local My Models shortcuts. Update Model List downloads a signed list only when clicked, for the exact Studio version and build. It does not send projects or model selections, download weights, or change existing projects. The delivery provider sees the version/build URL and ordinary connection information. Add Model remains available for compatible models outside the catalog.

Recommendations are shortcuts, not an allowlist. The development build can inspect a public Hugging Face repository ID or URL, or a local MLX text-model folder, without a catalog edit or application update. This is implemented selection behavior, not completed public support qualification. Model-picker spoken keyboard navigation was confirmed with the product owner. Lower-memory configurations remain unverified; unfamiliar-user feedback will be gathered during public beta, not required before launch.

In Baseline, open Choose Model, inspect the repository or choose a local folder, and review the model-specific source, license, and resource details. Repository inspection fetches metadata, not model weights; downloading needs separate confirmation. Missing license or resource information remains unknown. Once files are available, Test Compatibility runs a small synthetic training step and adapter reload before Use This Model becomes available. Passing this test is not a task-quality benchmark or a full-training memory guarantee.

• Reference model, not the only selectable model:mlx-community/Qwen3-4B-Instruct-2507-4bit at pinned revision50d427756c6b1b2fe0c0a10f67fbda1fc8e82c1b; model-card license metadata is Apache-2.0.
• Data: UTF-8 train.jsonl, valid.jsonl, and test.jsonl, with one object per line and string prompt and answer fields, or model-native messagesrecords. See data and response formats for training and evaluation requirements.
• Method: supervised LoRA/QLoRA through the bundled MLX runtime, with bounded Guided or Expert parameters.
• Evidence: dataset health, frozen baseline, candidate evaluation against the same test split, run state, source revisions, and artifact hashes.
• Export: a content-addressed LoRA adapter package with experiment, evaluation, and provenance evidence—not a fused model.

A project may remain open beside other projects, but only one operation runs inside each project. Concurrent training across projects has no public support guarantee and depends on available unified memory.

Use the model's own response format

Studio uses the selected model's supplied chat template. JSONL can contain prompt/answer pairs or native messages, including reasoning_content, tool_calls, tool results, and structured content where that template supports them. Optional tools and chat_template_kwargs are passed to the template; Studio does not execute tool calls or replace the template with its own.

Raw text records can be used for full-sequence training. Evaluation needs a prompt and expected answer, or a conversation ending in the expected assistant message; that final message is withheld from generation input. The method is full-sequence supervised fine-tuning with LoRA, not assistant-only loss, DPO, full-model tuning, or multimodal training.

Raw responses are preserved alongside the derived exact-JSON/text score. That score is not a general reasoning-quality or tool-use benchmark. A model completing a training run does not establish that it improved your task.

Set budgets within the selected model's capacity

• Recipe: training sequence budget, including the complete rendered example, prompt, answer, and template tokens.
• Baseline: input token budget, automatic by default.
• Baseline: response token budget, shared by baseline and candidate evaluation. Allow space for reasoning as well as the final answer.

The 512-token training and response defaults are starting points, not upper limits. Budgets follow the context declared by the pinned model configuration. Actual input plus requested response must fit that context. An unknown context is not unlimited support. Settings persist per project and selected model.

Oversized records stop with row and token-count guidance; Studio does not silently truncate training examples or input. Increase the relevant budget within model capacity if memory permits, or explicitly revise the dataset. Generation stops at the response budget even if the answer is unfinished. Declared context is not a RAM guarantee: long-context training can require far more memory than inference or the small compatibility test.

What the model tests establish

On 7 September 2026, all 17 pinned entries in the development catalog passed baseline, ten real training steps, comparison, adapter export, and fresh-process reload on the 128 GB Apple M4 Max test host. A separate audit verified finite adapter weights and learned values. Coverage includes Qwen, DeepSeek distilled models, Llama, Gemma, SmolLM, Mistral, and Liquid model families.

These are bounded technical workflow tests, not quality benchmarks, minimum-memory certification, maximum-context tests, or completed public release acceptance. New revisions still require their own checks.The 0.1.0 Beta 2 build 8 model catalog generation 3 is published for explicit-only refresh. The application update feed is published for build 8. Other version/build-specific catalogs remain unchanged.Catalog refresh availability is separate from the development build's offline model list.

Explicitly not supported in the first beta

• Intel Macs, Windows, Linux, iPhone, iPad, or virtualized macOS.
• Universal model compatibility, private or gated repository authentication, automatic weight-format conversion, or executing custom remote model code. Local import still requires a compatible model and appropriate usage rights.
• Generic document extraction, OCR, image/audio/video training, general chat, or agent workflows.
• Fused full-model, GGUF, Ollama, Core ML, or Hugging Face publishing export.
• Accounts, remote project storage, Multi-Mac execution, team roles, or sharing.
• Product analytics, automatic crash reports, background diagnostics, or a diagnostic upload service.
• A guarantee that fine-tuning improves quality or that an adapter works outside its recorded base-model/runtime boundary.

Start with the first incomplete stage

A support request should not become a data export

Include the Studio version/build, macOS version, chip and coarse memory tier, rounded free-disk range, workflow stage, exact visible error title/message and any displayed code, and concise reproduction steps. Crop or redact screenshots. Review every attachment.

Studio has no automatic report or upload path. The local engine's developer-facing doctor command can create a redacted bundle from explicitly selected JSON/YAML, but it is not a public in-app send feature. The first beta sends no product telemetry or automatic crash report.

Completed qualification and remaining release gates

Build 8 signing, notarization and retained-archive verification are complete. The exact artifact's third-party notice and native-source review is also complete. These checks do not establish public download availability, end-to-end update delivery or compatibility on untested Macs.

1. Complete signing-key custody and recovery verification for catalog publication. Any replacement artifact requires its own signing, notarization and archive verification.
2. Verify the final signed download, update feed, and release resources. Clean-device acceptance is deferred to beta follow-up, not claimed passed.
3. Keep hardware claims within measured evidence. Lower-memory Macs remain unverified beta configurations until directly tested.
4. Repeat the complete anonymous model download and network capture from the release account.
5. Resolve outstanding account, territorial, transfer and incident-response distribution evidence; publisher approval of the legal wording is already recorded. Recheck notices if the shipped artifact changes.