Run one private local AI prompt before paying for cloud inference
Use one Windows 11 model, a written prompt card and a small evidence log to decide whether local inference meets the job.
A repeatable local-prompt packet: one scoped prompt card, a chosen model record, three saved outputs, a privacy and hardware note, and a specific cloud-upgrade trigger.
Start with the output you need and the information that must stay local, not a model membership. It has not been run against Llama Windows or an external model, and the visual is a FreeStack concept rather than a provider screenshot or benchmark.
About this guide & its examples
This original FreeStack workflow uses a fictional product-summary task on Windows 11. Official sources checked 26 October 2026.
Illustrations are FreeStack concepts, not provider screenshots. Check the free limits below and test your own files before delivery.
Free boundaries before you start
What’s free, whether you need an account, and when we last checked.
Llama Windows · checked 2026-10-26
Free MIT-licensed Windows 11 tray application for local llama.cpp inference. The maintainer documents signed x64 and ARM64 MSIX releases, shared Hugging Face model storage, a local REST service and no telemetry. Model downloads and update checks use the network; model licences, disk space, hardware, electricity and output review remain separate.
The documented app and public local-inference route do not require an app account. Gated models and optional external services can require their own accounts.
Visual Studio Code · checked 2026-10-26
VS Code is free for private and commercial use on supported Windows, macOS and Linux systems. Core editing is not a hosted storage or AI allowance; extensions, Settings Sync, Copilot, remote services and deployment infrastructure have separate terms.
No Microsoft account is required for core local editing. Optional sync, marketplace or hosted services can require accounts.
Git · checked 2026-10-26
Free GPLv2 version-control software for local repositories, with no provider project quota. Remote hosting, backup, access control and large-file storage are separate services with their own limits.
No account is required for a local Git repository. Any remote host or sync service can require an account.
A small example to aim for
Summarise one product note
Keep three named facts. Use no more than 60 words. Add nothing the note does not support.
Make the boundary visible.
Buy only for the gap.
Fictional input · local model record · compared outputs · versioned decision
FreeStack workflow concept · not a benchmark, hands-on test or customer result.
Illustrative example · adapt it to your project.
Jump to a step
Write the job before choosing a model
In VS Code, create prompt-card.md for one harmless task: turn a fictional 180-word product note into a 60-word internal summary. Name the required facts, words to avoid and what a useful answer must contain. Keep personal, client and regulated information out of the experiment. Exclude web research, file access, tool calls and publishing so the first test has one observable boundary.
Choose a model that fits the machine
Install the signed Llama Windows release on Windows 11, inspect the memory estimate shown for a recommended model and read that model's own licence before downloading it. Start smaller than the available memory ceiling. Record the model ID, quantisation, context choice and source link. The MIT app licence does not grant rights to every model, training source or output.
Run the same prompt three times
Paste the same fictional note and prompt into the local overlay three times without adding new instructions. Save each response as plain text beside the prompt card and record the app and model versions. Compare only the written acceptance checks: word count, required facts and unsupported claims. Variation is evidence to review, not proof that one run is correct.
Check the privacy claim against the workflow
The maintainer states that prompts and inference stay local and that outbound requests cover model downloads, release checks and the local server user agent. Treat that as a current published claim, not an independent audit. Confirm your task needs no online source, remove sensitive text, note where logs and settings are stored, and decide whether the machine and account meet your organisation's controls.
Make the comparison inspectable
Put the prompt card, fictional source note, three responses and decision note in a local Git repository. Commit the untouched test pack, then commit the review. Do not place confidential material in a repository or remote host. Local Git gives named history; it is not an independent backup and does not validate model output.
Pay only for the missing requirement
Keep the local route when one Windows machine can produce acceptable drafts within the recorded time and hardware boundary. FreeStack editorial upgrade triggers: an approved hosted model, web-grounded research, shared access, managed uptime, larger context, supported governance, central audit retention or hardware you cannot justify owning. Buy the named capability only after the same prompt pack exposes the gap.
Your workflow, connected.
Follow the steps and carry the output into the next stage. These are suggested handoffs, not automatic integrations.
- 01Bound the prompt
VS Code core editing is free for private and commercial use. Extensions, sync and hosted AI services have separate terms.
Prompt card + acceptance checks - 02Record the model boundary
Llama Windows is MIT-licensed. It downloads models from Hugging Face and uses local hardware; model licences, disk space and electricity remain separate.
Model card + hardware note - 05Name the evidence
Git is free GPLv2 software. Remote hosting, backup, access control and large-file storage are separate.
Two review checkpoints
Provider information
Check current plans and documentation before committing to a workflow.
A couple of useful questions
Does this require an account?
The maintainer documents a signed Windows installer and local inference without an app account. Downloading a public model may not require sign-in, but gated models and external providers can. Record the model-specific requirement rather than assuming.
What stays free?
Llama Windows, VS Code core editing and Git provide ongoing-free software routes. Your Windows 11 device, storage, electricity, model licences, backup and any hosted inference remain separate.
Does local inference guarantee that nothing leaves the machine?
No blanket guarantee is made here. The current maintainer says prompts and inference stay local, while model downloads and update checks use outbound requests. Treat that as source evidence to compare with your own security requirements.
What does Save this toolkit keep?
The existing FreeStack flow saves the three catalogue tools to a private stack after sign-in. It does not copy prompts, models or outputs, run the app, create a repository or purchase inference.
What would you try?
Share your result or ask a specific question about this workflow.
Join the conversation