Local model readiness: choose before you download

By the end, you can

• Distinguish model weights, runtime, frontend, quantisation, context and agent harness before choosing a local stack (LA-7).
• Build a hardware and runtime preflight that records a Mac Studio's actual chip and unified memory instead of guessing from the product name (LA-7).
• Apply a download, supply-chain and privacy checklist, then justify a no-install local-model decision with evidence and explicit unknowns (LA-7).

Before you start

No model install or download is required for this lesson. This is the decision gate before the existing setup and first-run labs: you will write down what you have, what the model needs, which runtime you are actually choosing, and what evidence would make a download reasonable. If a value is unknown, leave it unknown and name how you will verify it. A product name such as Mac Studio is not a hardware specification.

Six layers people blur together

A local-model stack becomes much easier to reason about when each layer has one job:

Keep those names separate in your notes. If a recommendation says that a frontend supports a model, ask whether it means the frontend can send requests, the runtime can execute the weights, or the whole combination was tested on the target machine. Those are different claims.

• Model weights are the learned parameter files. They are the model artefact you obtain from a publisher or a permitted mirror; they are not the runtime or the chat application.
• Runtime or inference engine loads the weights and turns a request into generated tokens. It decides how the model is executed on the available CPU, GPU or accelerator.
• Frontend or client is the interface you use: a chat window, command-line client or API consumer. It can change without changing the model weights.
• Quantisation is a representation choice that reduces weight precision to trade memory and storage against retained quality and sometimes speed. It is applied to a model artefact; it is not a runtime and not a context window.
• Context is the token budget made available to a request. A model card's advertised maximum is not proof that a chosen runtime, machine or frontend will allocate that much context affordably.
• Agent harness is the surrounding code that may add a prompt loop, tools, memory, retries, logs and human approval. A capable model in a local runtime is not automatically a safe agent.

Qwen3.8-27B as a worked example

The official Qwen3.8-27B model card describes open weights under Apache-2.0, a dense 27B multimodal model and a native context length of 262,144 tokens. Its compatibility section publishes support paths for Transformers, vLLM, SGLang and TokenSpeed. The card also documents thinking control, including ways to disable or tune thinking for a request. These are useful facts to record, but they are not a complete local-readiness verdict.

The announcement and model card also contain vendor benchmark tables and release claims. Read those as evidence of what Qwen reported under its stated evaluation setup, not as independent proof of speed, quality, safety or fit on your machine. A 262,144-token native context length is not a promise that a local run should allocate 262,144 tokens. The actual context setting, runtime overhead, request contents and other applications all affect the decision.

This example is deliberately broader than one vendor. The lesson is teaching a repeatable preflight: identify the exact weights and license, check the runtime's current support, size the real machine, and test the job with evidence. Do not silently convert the Qwen card's published framework list into a claim of MLX or Ollama support for this model; this lesson has not verified either claim from a current primary source.

Mac Studio readiness worksheet

Use Apple menu → About This Mac, System Information or the matching Apple technical-specs page to fill the first two fields. Do not infer them from the words Mac Studio, the purchase year or somebody else's screenshot.

Unified memory is a shared pool, not a free model-only bucket. Before a real run, leave headroom for macOS, the runtime, the frontend, the request and the context-related working memory. Start from the exact model file or quantisation you intend to use and the runtime's current documentation; do not treat a rough parameter-count calculation as proof of fit. If the worksheet is incomplete, the safe result is not ready to download yet, not a guessed fit.

• Chip: ______________________________
• Unified memory: ____________________
• Free internal storage before any model work: ____________________
• Other work that must remain responsive while the model runs: ____________________
• Job to test first: ____________________
• Largest context the job actually needs: ____________________
• Runtime or runtimes being compared: ____________________

Choose a runtime without choosing a whole stack by accident

For Qwen3.8-27B, the official card publishes Transformers, vLLM, SGLang and TokenSpeed as supported framework paths. That tells you where to investigate first. It does not tell you that every path has the same hardware requirements, context behaviour, frontend integration, throughput or operational maturity on a Mac Studio.

Choose in this order: state the job and privacy boundary; record the actual chip and unified memory; identify the exact weight format and quantisation; compare a runtime's current support and memory controls; then choose a frontend and, only if needed, an agent harness. Keep the runtime and frontend as separate rows in your decision note. A frontend can be convenient while the underlying runtime is the part that determines whether the weights actually execute.

Download and privacy preflight

Treat model files, conversion scripts, launch commands and benchmark output as untrusted technical inputs until checked. Before downloading anything, record:

Do not paste a token, private document or learner record into a first-run prompt just because the model is local. Local execution can reduce an external data transfer, but it does not remove the need to inspect logs, frontend telemetry, network listeners, extensions and agent tools. The model license is not a privacy boundary, and a vendor benchmark is not a safety case.

• the canonical publisher and exact model-card URL;
• the owner, revision or release identifier you intend to use;
• the license and whether it permits your intended use;
• the exact format, quantisation and expected artifact size;
• a checksum or signed release record if the publisher supplies one;
• the runtime source and version that you plan to use;
• where the local cache will live and who can read it;
• whether the first test can run with no secrets, private learner data or customer data;
• how the runtime will be kept on the intended local network boundary; and
• what evidence would make you stop, delete the partial artefact or ask for human review.

No-install lab: make a go/no-go note

Use the worksheet and the Qwen3.8-27B example to make a short decision note. The point is not to force a download. A good beginner answer can be go, not yet, or no for this job, as long as the evidence and unknowns are visible.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The worksheet is made from ordinary text lines that can be copied into a note and completed with a keyboard. The practice exercise's model answer sits behind the existing native disclosure control. The knowledge check uses the existing native radio inputs and live status region, so incorrect answers can be retried and explanatory feedback is announced without a page reload.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

1. Qwen3.8-27B announcement — Alibaba Qwen. Vendor announcement used for the release context and headline model facts; its claims and benchmark figures are not independent proof.
2. Qwen3.8-27B model card — Qwen. Official model card used for the Apache-2.0 open-weights label, 27B dense multimodal description, 262,144 native context, published Transformers/vLLM/SGLang/TokenSpeed compatibility, thinking controls and vendor benchmark tables.
3. Mac Studio (2025) Tech Specs — Apple. Official lookup source for identifying a Mac Studio configuration; this lesson asks the learner to record their own chip and unified-memory values rather than assuming a configuration.