Portable Local Studio on Windows: setup, chat, images, speech and voice

1What you get

Portable Local Studio is a free, MIT-licensed app. Its browser interface has four workspaces: chat, images, transcription (speech to text) and voice (text to speech). It uses these open-source engines: llama.cpp for chat, stable-diffusion.cpp for images, whisper.cpp for transcription and Kokoro for voice.

• The app and the models are files in one folder. You can copy that folder to a USB drive and take it to another PC.
• The PC does all the calculation. Its CPU, GPU and memory set the speed. The USB drive only stores the files.
• The app needs the internet for setup and for model downloads. The code is made to run chat without the internet after that. Do the offline test in section 17 yourself before you depend on it.
• The app does not encrypt the drive. If you lose an unencrypted drive, anyone can read your chats.
• Each model has its own license. The app's MIT license does not cover the models.

Sections 6–9 cover chat. Section 10 covers images, section 11 transcription, section 12 voice, section 13 a voice loop that uses all three, and section 14 more examples. Section 15 shows how all files stay on the drive.

2Check your PC (2 minutes)

1. Press Ctrl+Shift+Esc to open Task Manager. Select Performance.
2. Write down the CPU name, the installed memory and each GPU name. Select a GPU to see its "Dedicated GPU memory". That is the VRAM.
3. Open Settings → System → About. Make sure that "System type" shows 64-bit operating system, x64-based processor.
4. Make sure that you have at least 15 GB of free space on the target drive. The test used 8.1 GB for the app, two chat models, one image model, one speech model and one voice model.

3Choose a folder or a USB drive

Install on the internal disk first. When chat works there, copy the full folder to the USB drive. A problem is easier to find on the internal disk, because a slow or faulty drive cannot cause it.

How the test simulated a 128 GB USB drive: The test laptop had no USB drive connected, so we copied the app to C:\AI\usb128. We then ran subst U: C:\AI\usb128 to make that folder appear as drive U:. The app ran only from U:\. This proves that the app works from any drive letter. It does not test USB speed or exFAT. TESTED

Step 4Download and extract the app

1. Open the project's GitHub page: github.com/techjarves/Portable-Local-Studio (MIT license, by Tech Jarves). The old name in the app's code, github.com/techjarves/Uncensored-AI-Studio, redirects to it. CHECKED
2. Click the green Code button. Then click Download ZIP.
3. Right-click the ZIP file. Click Extract All. Extract to your target, for example C:\AI\ or E:\.
4. Rename Portable-Local-Studio-main to Portable-Local-Studio if you want. This step is optional.
5. Open the extracted folder. Make sure that it holds windows.bat, linux.sh, mac.sh, app and scripts.

Step 5Run the first-time setup

To see which program uses a port, run this command in PowerShell:

Get-NetTCPConnection -LocalPort 8080,10086 -State Listen -ErrorAction SilentlyContinue | ForEach-Object { Get-Process -Id $_.OwningProcess }

No output means that both ports are free.

1. Make sure that the PC is online.
2. Double-click windows.bat. A black terminal window opens.
3. Read the banner. It says First-Time Setup and gives the reason, for example "Portable Node.js is missing."
4. Press any key to continue. Press Ctrl+C instead to cancel.
5. Wait. Setup has 8 steps. On the test laptop, the first run took 16.8 minutes. Your network speed sets most of this time.
6. If Windows asks for permission to install the Microsoft Visual C++ Redistributable (x64), check that the publisher is Microsoft. Then click Yes. On a work PC, you may need your administrator.
7. When setup ends, the browser opens at http://localhost:1420.

Step 6Download the starter model (Qwen2.5 Coder 0.5B)

The starter model is small (491,400,064 bytes; the app shows "468.6 MB"). It loads fast and proves that the chat engine works. It is too small to show what local AI can do. Section 9 adds a better model.

1. In the sidebar, click Models.
2. Click the Text tab. The default tab is Image, so make sure that you change it.
3. Under Library → Recommended text models from Hugging Face, find Qwen2.5 Coder 0.5B Instruct Q4_K_M (~491 MB).
4. Click Download on that card.
5. Wait until the model shows under On this computer. The download took 11.9 s on the test connection. TESTED (through the app's API, which the button uses)

https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct-GGUF/resolve/main/qwen2.5-coder-0.5b-instruct-q4_k_m.gguf

The model file goes to app\llm-models\. You can also copy a .gguf file into that folder yourself. Text models go in app\llm-models\, and image models go in app\models\. Do not mix them.

Step 7Have your first chat

1. In the sidebar, click New chat.
2. Click Select model near the bottom-right corner of the message box.
3. Click qwen2.5-coder-0.5b-instruct-q4_k_m. Wait until the top bar shows a green dot and the model name. On the test laptop, the model loaded in 1.6–2.1 s.
4. Make sure that Web is off. Then the chat stays on your PC.
5. Type this test message and press Enter:

Extract exactly three fields: name, date, amount.
Text: Alice paid $37.50 on 2026-10-31.
Do not add any other information.

Expected answer (TESTED, three times):

Alice, 2026-10-31, $37.50

Compare each field with the message. All three must be correct. This test proves that the chat engine works. It does not prove that the model reasons well.

If you see the correct answer, the app works on your PC. For a model that writes better answers, go to sections 8 and 9.

Step 8Set the KV cache before you load a second model

The KV cache is the model's working memory for the conversation. The app can store it at three precisions: Q4_0 (smallest), Q8_0 and F16 (largest). In the test, Q4_0 made the Qwen2.5 1.5B model write nonsense. Q8_0 and F16 gave correct answers.

1. In the sidebar, click Settings.
2. Click the Text generation tab.
3. Under Performance → Profile, click Balanced.
4. Under KV cache, click Q8_0.
5. If an "Unsaved text settings" bar appears, save the changes in it.

Step 9Add a better model: Qwen2.5 1.5B Instruct

This model is a general chat model of 1,117,320,736 bytes (the app shows "1.0 GB"). Its answers are better than the 0.5B coder model's answers. It is not in the app's library, so you download it from a link.

1. Click Models. Click the Text tab.
2. Click + Add model at the top-right corner.
3. Select Download from a link.
4. Paste this URL into the box:

https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/resolve/main/qwen2.5-1.5b-instruct-q4_k_m.gguf

5. Click Download. On the test connection, the download ran at 69.7 MB/s. TESTED
6. Wait until On this computer shows 2 models.

7. Click New chat.
8. Click the power icon beside the model name in the top bar. Or open the model menu and click Unload model. The top bar must show No model loaded.
9. Click Select model. Click qwen2.5-1.5b-instruct-q4_k_m.
10. Wait for the green dot. Then ask a question, for example "In two sentences, explain what a USB flash drive is."

Step 10Make images (DreamShaper 8)

The Images workspace uses stable-diffusion.cpp. DreamShaper 8 is a Stable Diffusion 1.5 model of 2,132,625,894 bytes (the app shows "~2.1 GB"). It makes 512×512 images and fits in 4 GB of VRAM. It downloads from Hugging Face without a login.

Download the image model

1. Click Models. The Image tab opens by default.
2. Under Library → SD 1.5 — fast, low memory, find DreamShaper 8 (~2.1 GB · 512x512).
3. Click Download on that card. The dialog "Download "DreamShaper 8" (~2.1 GB)?" opens.
4. Click Download in the dialog.
5. Wait until the model shows under On this computer. On the test connection, the download ran at about 57 MB/s and took about 40 s. TESTED

Make your first image: a product illustration

1. In the sidebar, click Images.
2. Click Select model at the bottom-right corner. Click DreamShaper_8_pruned.
3. If a chat model is loaded, the dialog "Unload the chat model?" opens. Click Unload and load.
4. Wait until the top bar shows a green dot, Image and DreamShaper_8_pruned.
5. Make sure that the bar under the prompt shows 512×512 and 20 steps. The app set these values when the model loaded. Before the load, it showed 1024×1024 and 4 steps. TESTED
6. Click More. Make sure that CFG scale is 7. Type this in Negative prompt (what the image must not show):

blurry, low quality, text, watermark, deformed

7. Press Esc to close More. Type this prompt in the prompt box:

product illustration of a matte black reusable water bottle on a light wooden table, soft studio lighting, clean background, high detail

8. Press Enter, or click the arrow button.
9. Wait for the image. On the test laptop, it took 34.3 s: 27.2 s for the 20 steps (about 1 s per step after a 7.9 s first step) and 4.6 s to decode the image. TESTED

The app saves each image as a PNG file and a JSON file with the prompt, seed, steps and size in app\outputs\. The test image was 381,033 bytes. To make the same image again, use the same prompt, model, size, steps and seed. TESTED (files)

Step 11Transcribe speech to text (Whisper)

The Transcribe workspace uses whisper.cpp on the CPU. It changes a recording into text. Whisper Base English is 147,964,211 bytes (the app shows "142 MB").

Download the speech model

1. Click Models. Click the Speech tab.
2. Find Whisper Base English (142 MB · Whisper.cpp). Click Download.
3. In the dialog "Download "Whisper Base English" (142 MB)?", click Download.
4. Wait until the model shows under On this computer. The file goes to app\speech-models\ggml-base.en.bin. TESTED

Transcribe a voice memo

1. In the sidebar, click Transcribe. Make sure that the top-right menu shows Whisper Base English.
2. Click Upload audio. Select an audio file on the drive. The app accepts WAV, MP3, M4A, OGG, FLAC and WebM. The browser converts the audio to 16 kHz mono before it sends it to the engine.
3. Click Transcribe beside the file.
4. Wait for the text under Transcript. Read it and correct it.
5. Click Copy transcript to copy the text, for example for a chat.

Test file: a 14.4 s voice memo made in the Voice workspace (section 12). The transcription took 9.2 s on the first run, with the engine on 8 CPU threads. An 8.0 s question took 2.0 s. TESTED

All of the words were correct. Whisper wrote the numbers as digits ("200", "3P. M.") and removed one comma.

The app saves each transcript as a TXT file and a JSON file in app\transcriptions\. TESTED

• Record with the microphone: click the microphone button (Start recording). The browser asks for permission to use the microphone. Click Allow. NOT TESTED
• Other languages: "English" models cannot transcribe other languages. The app then says: "Download a multilingual Whisper model (for example Whisper Base Multilingual)" (SpeechTranscriber.jsx:407). Use Spoken language and Translate to English with a multilingual model. NOT TESTED

Step 12Turn text into speech (Kokoro)

The Voice workspace uses Kokoro 82M on the CPU. It reads text aloud and saves a 24 kHz WAV file.

Install the voice model

1. Click Models. Click the Voice tab.
2. Find Kokoro 82M ONNX Q8. Click Download.
3. The dialog says "Kokoro model files are cached on first generation." Click Install.

Read a short script aloud

1. In the sidebar, click Voice. Make sure that the top-right menu shows Kokoro 82M ONNX Q8.
2. Type the text in the box at the bottom. The limit is 5,000 characters.
3. Click the voice button (default Heart). Select a voice.
4. Click the speed button (default 1×) if you want a different speed.
5. Click Speak, or press Enter.
6. Wait for the clip. It plays automatically and marks each word as it plays.

Test text:

Voice memo for Monday. Call the print shop about the two hundred flyers, they are due on Thursday. Ask Dana for the budget sheet before noon. The team meeting moves to three p.m. on Wednesday.

Result: a 14.4 s clip (690,524 bytes). The first Speak took about 14 s from the click to the saved file. TESTED

The app saves each clip as a WAV file and a JSON file with the text, voice and timing in app\tts-outputs\. TESTED

Write times as "3 PM", not "three p.m.". The app cut "three p.m." into two parts ("three p." and "m. on Wednesday"), because it splits the text at each period. TESTED The microphone button in the Voice box (Dictate) types your speech into the box. NOT TESTED

Step 13A "Jarvis" loop: ask by voice, hear the answer

The app has no automatic voice-assistant mode. The chat page only has shortcut buttons to the other workspaces (TextChat.jsx:917-918). You connect the three workspaces yourself with copy and paste. The chat model, Whisper and Kokoro can run at the same time: Whisper and Kokoro use the CPU, and only the chat model uses the GPU. TESTED

1. Load the chat model (section 9). Keep the KV cache at Q8_0.
2. Click Transcribe. Record your question, or upload an audio file with the question.
3. Click Transcribe. Click Copy transcript.
4. Click New chat. Paste the question with Ctrl+V. Press Enter.
5. Select the answer text. Press Ctrl+C.
6. Click Voice. Paste the answer. Press Enter. The answer plays aloud.

All four hops worked. TESTED The question asked for three ways, and the 1.5B model gave about two. Small models do not always follow counts.

Step 14More examples

A wide blog header image

1. Click Images. Load DreamShaper 8 (section 10).
2. Click the size button. Select 16:9 · 768×432.
3. Type this prompt and press Enter:

wide blog header, cozy home office desk with a laptop and a cup of coffee at sunrise, warm light, plants, shallow depth of field, digital painting

Result: 768×432 · 20 steps · 29.0 s · 709,066 bytes. The size menu also has 4:3 (640×480) and 9:16 (432×768). TESTED The second row of sizes (1024×1024 to 1216×684) matches SDXL sizes. The test did not use it.

Meeting notes: from a voice memo to a to-do list

1. Transcribe the memo (section 11). Click Copy transcript.
2. Load the chat model. If the image model is loaded, the dialog "Unload the image model?" opens. Click Unload and load.
3. Click New chat. Type the instruction below, paste the transcript after it and press Enter.

Turn this voice memo transcript into a to-do list with one line per task and its deadline. Transcript:

Result in 999 ms (38 tokens): TESTED

Call print shop about flyers due on Thursday by 4PM.
Request budget sheet from Dana before noon.
Move team meeting to 3PM on Wednesday.

Compare two voices

1. Click Voice. Type a short text.
2. Select Heart (en-us). Click Speak.
3. Type the same text again. Select George (en-gb). Click Speak.
4. Click the play button on each clip in the list. Keep the voice that you prefer.

The test made one clip with Heart and one with George, with different texts. Both clips are in app\tts-outputs\, and the voice ID is in each file name (af_heart, bm_george). TESTED

Step 15Keep everything on the 128 GB drive

All parts of the studio fit on a 128 GB drive with a large margin. After all tests in this manual, the project folder held 8,096,075,478 bytes (8.10 GB, 7.54 GiB). A 128 GB drive has about 119.2 GiB of usable space, so about 111.7 GiB stays free. TESTED

With the free space, you can add, for example, an SDXL Lightning model (6.9–7.1 GB) or a 7B chat model (4.7 GB). Count about 0.4–0.7 MB for each image and about 48 KB for each second of speech.

How we checked for files outside the drive

We recorded the time before the tests in this manual (1 Nov 2026, 07:09). After the tests, we listed each file larger than 100 KB that changed after that time. We searched these folders: %USERPROFILE%\.cache, %LOCALAPPDATA%\npm-cache, %APPDATA%\NVIDIA, %LOCALAPPDATA%\NVIDIA, %TEMP% and Downloads. TESTED

• No model, cache or output of the studio was outside the drive. The Hugging Face cache (%USERPROFILE%\.cache\huggingface) and the npm cache did not change. The CUDA cache (%APPDATA%\NVIDIA\ComputeCache) did not change.
• Two writes came from the browser, not from the studio: the browser profile (page cache, history) and the NVIDIA shader cache %LOCALAPPDATA%\NVIDIA\DXCache.

Rules to keep the studio on the drive

1. Run the first setup with the npm cache on the drive (section 19, "Fix for the npm ENOENT error"). Without it, npm writes to %LOCALAPPDATA%\npm-cache on drive C:.
2. Use the files that the app already saved: app\outputs, app\tts-outputs and app\transcriptions. The Download, TXT, JSON and Download WAV buttons save a second copy through the browser, normally to C:\Users\<you>\Downloads.
3. If you use those buttons, set the browser to ask where to save each file. Then select a folder on the drive.
4. Keep the audio files that you transcribe on the drive. While it works, the app writes a temporary copy of each upload into app\transcriptions.
5. Use a private (Incognito or InPrivate) browser window for the studio if you want less browser history on the PC. The browser profile is outside the app's control.

Two things outside the app can write to drive C: on a new PC. Setup installs the Visual C++ runtime into Windows if it is missing (setup.ps1:213-239). The GPU driver can keep shader caches in %LOCALAPPDATA%\NVIDIA. Neither holds your models, prompts or outputs.

Step 16Stop, restart and move the app

Stop FROM CODE

1. Click into the terminal window of windows.bat.
2. Press Ctrl+C. If Windows asks "Terminate batch job (Y/N)?", type Y and press Enter.
3. Wait until the window closes or shows a prompt. Then all services are stopped.

Start again

1. Double-click windows.bat. Setup does not run again if all parts are present.
2. Open the address from the terminal's "Web UI:" line.
3. Select a model in the chat. Your earlier chats are listed under Chats in the sidebar.

A full restart of the test install loaded the model in 1.6 s and gave the same correct answer with no new downloads. TESTED

Move the app to a USB drive

1. Stop the app.
2. Copy the whole project folder to the drive: engines, runtimes, models, settings and chats. The original ZIP alone does not contain the prepared engines.
3. Start the copy from the drive. Make sure that an earlier chat is visible and that a new answer works.
4. Keep the internal copy until the drive copy works.
5. When you finish a session, stop the app first. Then use Safely Remove Hardware to eject the drive. The app reads models directly from the drive while it runs.

Step 17Do an offline test NOT TESTED

A model that works while you are online does not prove that the next offline launch works. Do this test before you travel or depend on the app.

1. While you are online, download every model that you will need: chat, image, Whisper and Kokoro.
2. Stop the app with Ctrl+C.
3. Disconnect Wi-Fi, Ethernet and phone tethering.
4. Double-click windows.bat. At each launch, it checks for a llama.cpp update. Offline, it shows "[WARN] Could not update the llama.cpp backend (offline?)" and continues (update-llama.ps1:31-32). FROM CODE
5. Load a model in a new chat. Send the test message from section 7.
6. Pass: you get a new, correct answer. An old chat in the history does not count.
7. Make one new image, one transcript and one voice clip in the same way.

18. AMD, Intel and CPU-only PCs

Setup detects the GPU vendor and installs the matching chat engine (setup-llama.ps1:190-236). It always installs the Vulkan and CPU engines too. In Settings → Text generation → Text backend, you can select any installed engine.

A CPU-only PC or a PC with a small GPU also goes into the "Potato" tier. Set the KV cache to Q8_0 as in section 8.

19. Windows gotchas and fixes

When something fails, read the terminal window first. The line before "server crashed" usually names the cause.

Fix for the npm ENOENT error

1. Open the project folder in File Explorer.
2. Click the address bar. Type cmd and press Enter. A Command Prompt opens in that folder.
3. Run these two commands:

set npm_config_cache=%CD%\app\tools\npm-cache
windows.bat

Setup skips the parts that are already done. On the test laptop, this second run took 0.3 minutes and ended with exit code 0. TESTED

20. Repair and advanced route

Use these PowerShell commands to repair a broken install or to test the app without the browser. Run them in the project folder while you are online.

Repair the installation

Stop the app first. Then run:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\setup\setup.ps1

This is the same setup that windows.bat runs. It keeps your models and chats. TESTED The project also has scripts\reset\reset.ps1, which removes caches. Make a backup of the whole folder before you use it. FROM CODE

Test chat without the browser

Start the app with windows.bat. Then run this in a second PowerShell window. Change 1420 if the terminal shows a different port.

$base = "http://127.0.0.1:1420"
Invoke-RestMethod "$base/api/llm/status" | Select-Object ready, backendMode
Invoke-RestMethod -Method Post "$base/api/llm/stop"
$load = @{ model = "qwen2.5-1.5b-instruct-q4_k_m.gguf"; cacheTypeK = "q8_0"; cacheTypeV = "q8_0" } | ConvertTo-Json
Invoke-RestMethod -Method Post "$base/api/llm/start" -ContentType "application/json" -Body $load | Select-Object ok, ready
$msg = @{ stream = $false; temperature = 0; max_tokens = 60; messages = @(@{ role = "user"; content = "Extract exactly three fields: name, date, amount. Text: Alice paid `$37.50 on 2026-10-31. Do not add any other information." }) } | ConvertTo-Json -Depth 4
(Invoke-RestMethod -Method Post "$base/api/llm/chat" -ContentType "application/json" -Body $msg).choices[0].message.content

Expected output: the three fields Name: Alice, Date: 2026-10-31 and Amount: $37.50. The 1.5B model can add a short heading or a list, so the format can be different. TESTED (these exact commands)

21. Measured results on the test laptop

The 0.5B model also wrote that a USB drive works "without physically inserting it into the computer". That is wrong. Small models make factual mistakes, so check their answers.

22. What this test did not cover

• Real USB speed. Files came from the internal SSD through subst. A USB 3.0 flash drive at 100–200 MB/s needs about 2.5–5 s just to read the 491 MB model, and about 5.6–11 s for the 1.12 GB model. After the model loads, answer speed should stay the same, because the model then runs from RAM or VRAM.
• exFAT. The test folder was on NTFS.
• The offline test (section 17). Without administrator rights, we could not switch off the network.
• AMD, Intel Arc and CPU-only PCs as the main hardware. Only detection and CPU speed were measured.
• The Download button for the 0.5B card. The test started this download through the app's API (/api/llm/download-model), which the button also uses. The card then showed "Downloaded".
• The Visual C++ install prompt. The runtime was already on the test laptop.
• SmartScreen and antivirus prompts. None appeared.
• The microphone (Start recording, Dictate). The test used audio files that the Voice workspace made.
• Other languages and Translate to English in Transcribe.
• SDXL models, "Use as base" (image to image), file attachments and web search.
• The upload dialog with a file on U:\. The test tool could upload only from its work folder, so we copied the two test WAV files there and deleted them after the test. A normal user selects the file on the drive.

23. Versions and code references