zorlo Download

Version 0.1 Beta · Qwen3-30B-A3B · Windows 10 and 11 · Free

A 30B model.
An 8 GB PC.

Zorlo runs Qwen3-30B-A3B, a 30-billion-parameter AI model, on a Windows PC with 8 GB of memory. The model runs entirely on your computer, so your conversations and files stay with you. Beyond chat, Zorlo works as a coding agent: it writes code, builds websites, reads your documents and runs commands in a folder you choose.

Download for Windows

5 MB · Windows 10 or 11, 64-bit
Public beta · no installer, no account

Model
Qwen3-30B-A3B
Model size
18.6 GB
Memory needed
8 GB
Writing speed
2–3 words/s
vs llama.cpp
2.4× faster
Runs
on your PC

How it works

A model larger than your memory, loaded one piece at a time.

  1. 01

    The model is made of experts.

    Qwen3-30B-A3B has 128 small expert networks in each of its 48 layers. For every token it writes, it consults only 8 of them.

  2. 02

    Zorlo loads only those 8.

    The full model stays on your SSD. For each token, Zorlo reads in just the experts that token needs (orange squares).

  3. 03

    Frequently used experts stay in memory.

    Experts that are needed again and again are kept in memory, so they are read from disk only once (outlined squares, orange flashes).

  4. 04

    The result: a 30B model on an 8 GB PC.

    The model is 18.6 GB; Zorlo keeps less than 1 GB of it in memory at all times. In the default mode the output is identical to running the full model with llama.cpp.

128 experts on the disk; for each token 8 light up, the ones not already in memory are loaded, then the processor writes the token.
A simplified view of one layer. The reuse rate shown is taken from real runs.

2.4×

the speed of llama.cpp with the same model on the same 8 GB PC

5.47

tokens per second in fast mode (a token is about ¾ of a word)

0.4 s

to resume a long conversation, even after a restart (69 s the first time)

1280/1280

tokens identical to llama.cpp across 20 test prompts, in the default mode

Performance

Measured, not estimated.

All figures are for Qwen3-30B-A3B in 4-bit form (Q4_K_M, 18.6 GB), measured on a reference PC with 8 GB of memory and an NVMe SSD, running Windows 11. Speed is in tokens per second. The comparison is llama.cpp with its default settings. Your results depend mainly on the speed of your SSD and on how much memory is free.

Writing speed

Tokens per second, higher is better

Decode speed in tokens per second
llama.cpp1.43
Zorlo, 256 MB of memory (what a 4 GB PC would have available)2.98
Zorlo, 1.2 GB of free memory3.47

Same model, same output. Zorlo knows in advance which experts each token needs and reads exactly those; llama.cpp relies on Windows to page the model in.

Speed modes

Tokens per second with little free memory, and how closely the output matches

Speed modes: tokens per second and share of identical tokens
exact default · 100% identical2.25
balanced 98% identical2.53
fast 95% identical5.47

Fast mode uses a compact copy of rarely used experts, so an occasional word can differ. Exact mode never changes the output.

Long prompts, cached

Seconds before the first token of a long request, lower is better

Prompt processing time in seconds
first request69.0 s
same prompt again4.3 s
after a restart0.4 s

Zorlo saves the work it has already done on a prompt to disk, so agents that resend the same long instructions continue from where they left off.

Why it fits

Gigabytes

Model size versus PC memory versus what Zorlo keeps in memory
model on disk18.6 GB
memory in an 8 GB PC8 GB
what Zorlo always keeps in memory0.82 GB

Any other free memory is used to keep popular experts at hand, and is released as soon as another application needs it.

Get started

Up and running in three steps.

No installer and no account. You need Windows 10 or 11, 8 GB of memory, about 19 GB free on an SSD, and an internet connection for the one-time model download. Full requirements.

  1. 01

    Download and extract

    Download for Windows5 MB

    In your Downloads folder, right-click zorlo-v0.1-windows-x64, choose Extract All…, then Extract.

  2. 02

    Open Zorlo Desktop

    Double-click Zorlo Desktop in the extracted folder. On first start, choose a model; Qwen3 30B is recommended. The model is downloaded once (about 18.6 GB), with a progress bar and automatic resume if the connection drops. If the model is already on your PC, Zorlo finds it and uses it from where it is; you can also point to its folder with Already have this model?

    If Windows shows "Windows protected your PC", select More info, then Run anyway. Windows shows this for new applications that are not yet widely installed.

  3. 03

    Start chatting

    Once the model has loaded (about a minute), the chat opens in the same window. From then on, Zorlo Desktop starts the model directly and keeps it up to date automatically. Zorlo also adds itself to the Start menu: next time, press the Windows key, type Zorlo and press Enter.

    When you close the window, Zorlo asks whether to keep running in the background, with the model ready for your next question, or to stop and free the memory. The Zorlo icon next to the clock lets you reopen Zorlo, switch models, start Zorlo with Windows, or Stop Zorlo and quit.

Using a terminal instead

Zorlo can also be started and stopped from Windows PowerShell. The chat then opens in your web browser. After Zorlo has started once, by either method, the zorlo command works in any new PowerShell window, from any folder.

  1. The first time: open PowerShell in the Zorlo folder. Open the extracted folder in File Explorer, click the address bar, type powershell and press Enter. If you already opened Zorlo Desktop once, skip this: any PowerShell window works.
  2. Start Zorlo. The first time, this lets you choose and download a model; after that it starts the model directly. When the model is ready, the chat opens in your browser. 
    .\zorlo start
    From then on, in any new PowerShell window: 
    zorlo start
    If PowerShell says 'zorlo' is not recognized, Zorlo has not been started yet on this PC, or the window was open before it was: open a new PowerShell window in the Zorlo folder and use .\zorlo start.
  3. Open the chat. If the browser does not open by itself, or you closed it, go to http://127.0.0.1:8080, or run: 
    start http://127.0.0.1:8080
    Keep the PowerShell window open while you use Zorlo. Closing the browser does not stop Zorlo.
  4. Stop Zorlo. Click the PowerShell window and press Ctrl+C, or close the window. The model is unloaded and its memory is released.
  5. Start it again later. Open PowerShell and type zorlo start. Your model is not downloaded again. (To remove the zorlo command from other folders, run zorlo path off.)
  6. If Zorlo reports that it is already running. Another copy is still open, for example in the background or in another window. Stop all copies with: 
    Stop-Process -Name zorlo
    Alternatively, open Task Manager, select zorlo.exe and choose End task.
Advanced: download or convert the model manually

zorlo pull resumes interrupted downloads, verifies every file and later fetches only files that changed. Add --lowbit for the optional fast mode (an extra 9.5 GB). Use the folder below so Zorlo Desktop finds the model.

.\zorlo pull Haidershahi1/qwen3-30b-a3b-zorlo models\qwen3-30b-a3b

To build the model files yourself, place Qwen3-30B-A3B-Q4_K_M.gguf (about 18.6 GB, from Hugging Face) next to zorlo.exe and run once:

.\zorlo pack Qwen3-30B-A3B-Q4_K_M.gguf models\qwen3-30b-a3b

To start a downloaded model directly, without the model menu or the update check:

.\zorlo serve models\qwen3-30b-a3b -c 12288

zorlo-v0.1-windows-x64.zip · 5,452,347 bytes · SHA-256 4fa88cbfb1a8fc078602cd884d18a5e98c6dbdd0d6f6b89ff61aa585993b3e55

Features

A complete assistant, running on your PC.

The Zorlo app in Agent mode: tool cards for writing ember-site/style.css (with an Undo button) and preview_url, the agent's summary of the coffee-shop website it built, and the Preview panel showing the finished Ember Beans page with its menu and opening hours.
Zorlo in Agent mode with Qwen3-30B-A3B. Asked for a one-page coffee-shop website, it planned the work, wrote index.html and style.css with the user's approval, and displayed the result in the Preview panel.

  • Coding agent

    Choose a project folder and describe the task. Zorlo plans, reads and edits files, runs commands, searches the web and previews websites as it builds them. Every file change can be undone, and Zorlo asks before any action you have not allowed in Settings. Plan mode proposes the steps first and changes nothing.

  • Documents, spreadsheets and images

    Attach PDF, Word, Excel and PowerPoint files, images, code or text, and ask about them. Zorlo extracts the text, reads scanned pages and the text in images with Windows' built-in text recognition, and gives the model the length of videos and audio. Qwen3-30B-A3B is a text model: it reads the text of an image, not the image itself.

  • Connectors and skills

    Extend the agent with connectors (MCP servers) for GitHub, Notion, Linear, a real browser, databases, search and more, added with one click from the built-in catalog, or by asking in a chat. Skills add expert instructions for tasks such as creating Word, Excel, PowerPoint and PDF files, and install from GitHub in one step.

  • Private by design

    With Zorlo's own model, everything runs on your PC; conversations and files are not sent to any AI service. Zorlo uses the internet only to download models, and for web searches and connectors that you enable. Each chat is kept with its project folder, and several chats can run at once.

  • Bring your own model

    Run the same agent, with all its tools, on a cloud model. In Settings → Model, choose OpenAI, Anthropic, Google Gemini, DeepSeek, OpenRouter, Groq, Mistral, xAI or Together AI, or any OpenAI-compatible server, including Ollama and LM Studio on your PC. Enter your API key and a model name. The key is stored on your PC and sent only to the provider it belongs to; conversations then go to that provider.

  • OpenAI-compatible API

    Use Zorlo from your own code or tools: point any OpenAI client at http://127.0.0.1:8080/v1 with any API key. Streaming, tool calls and reasoning are supported.

    from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="zorlo")
r = client.chat.completions.create(model="zorlo",
    messages=[{"role": "user", "content": "Hello"}])
print(r.choices[0].message.content)

  • Anthropic API, for Claude Code

    Zorlo also implements the Anthropic Messages API, so Claude Code can run on the local model. Start Zorlo with a larger context for Claude Code's instructions.

    .\zorlo serve models\qwen3-30b-a3b -c 20480
set ANTHROPIC_BASE_URL=http://127.0.0.1:8080
set ANTHROPIC_AUTH_TOKEN=zorlo
set ANTHROPIC_MODEL=zorlo
claude

  • Speed modes

    Choose exact (default, identical to llama.cpp), balanced or fast in the app, as a server default with --mode, or per request with a "mode" field. Fast mode requires the compact expert copy: zorlo pull … --lowbit.

    # server default for every client
.\zorlo serve models\qwen3-30b-a3b -c 12288 --mode fast

# per request, from your own code
client.chat.completions.create(model="zorlo", messages=msgs,
    extra_body={"mode": "exact"})

In practice

Results from real runs.

Recorded on the reference PC with Qwen3-30B-A3B. Speeds vary by about 20% between runs.

tool calling · three modes

A weather request for Paris

→ get_weather {"city": "Paris"}

fast 5.47 tok/s · balanced 3.13 · exact 2.70

All three modes returned the same valid tool call. In exact mode the response matches llama-server byte for byte.

code edit

Renaming a variable in a 55-line Python file

draft tokens accepted 243 / 243

2.70 → 3.26 tok/s · output identical

When the answer repeats earlier text, Zorlo drafts several tokens ahead from that text and verifies them in a single pass.

agent prompt · restart

The same long prompt after a restart

prompt processing 69.0 s → 0.4 s

cached state read from disk in 0.12 s · same reply

A system prompt with 60 rules and two tools. After a restart, the saved state loads from disk and the reply is identical to the first run.

documents · images

A question about a photo and a spreadsheet

ticket.png + scores.csv attached

"Your train is on Platform 6, and Ada scored higher."

Zorlo read the text in the photo with Windows text recognition and the table from the CSV file, and the model answered from both.

System requirements

What you need for Qwen3-30B-A3B.

These requirements apply to Qwen3-30B-A3B by the Qwen team, the model offered in Zorlo 0.1, in 4-bit form (Q4_K_M). It has 30.5 billion parameters, of which 3.3 billion are active for each token. Future models will list their own requirements.

Operating system
Windows 10 or 11, 64-bit.
Processor
A processor with AVX2: Intel Core 4th generation (2013) or newer, or AMD Ryzen or newer. Processors with AVX-512 VNNI use a faster path automatically.
Memory
8 GB recommended. Zorlo reserves about 2 GB (0.8 GB of model data plus the conversation's working memory) and uses any additional free memory to answer faster.
Storage
19 GB free on an SSD, or 28 GB with the optional fast mode. An NVMe SSD is recommended; a SATA SSD works at lower speed. Hard disks are not supported.
Expected speed
About 2.2 to 3.5 tokens per second (2 to 3 words), up to 5.5 in fast mode, depending on your SSD and free memory.

Model-specific figures. All numbers on this page were measured with Qwen3-30B-A3B. It runs on an 8 GB PC because it is a Mixture-of-Experts model that uses only a small part of its weights for each token. Larger models, or models that use all of their weights for every token, need more memory and storage and run slower. Each new model will be published with its own tested requirements, and Zorlo Desktop shows the download size and memory requirement before you choose a model. Zorlo 0.1 is available for Windows.

How do I check my PC?

Open Settings → System → About to see your processor and installed memory (RAM), and Settings → System → Storage for free space. Any Intel Core processor from 2015 or later, or any AMD Ryzen processor, is supported.

What does "beta" mean?

Zorlo 0.1 is a public beta: it is complete and tested, but has been measured on a limited range of PCs so far. Results on your PC may differ, especially with less free memory or a slower SSD. When a new version is available, Zorlo tells you when it starts.

Is Zorlo free?

Yes, for personal and commercial use. Zorlo is licensed under the Functional Source License (FSL-1.1-MIT): you may use, modify and share it for any purpose except offering a competing product or service. Each release becomes available under the MIT license two years after publication.

Is my data private?

Yes. The model runs entirely on your PC, and your conversations and files are not sent to any AI service. The internet is used to download models and, if you allow it, for web searches (DuckDuckGo, or Brave Search with your own key) and connectors you add. Web access can be turned off in Settings. If you choose a cloud provider in Settings → Model, conversations are sent to that provider under its terms.

Are the answers as good as the full model?

Yes. In the default exact mode the output is identical to running the full model with llama.cpp; in our tests all 1,280 tokens across 20 prompts matched. The optional faster modes may change an occasional word.

Can the model see pictures?

Qwen3-30B-A3B is a text model. Zorlo reads the text in pictures and scanned documents with Windows' built-in text recognition and passes it to the model, so it can answer questions about receipts, screenshots and scans, but it cannot describe a photo's visual content.

How do I stop or uninstall Zorlo?

Close the Zorlo window and choose Stop Zorlo, or right-click the Zorlo icon next to the clock and choose Stop Zorlo and quit. To uninstall, delete the Zorlo folder. Zorlo does not install anything else; if you turned on Start with Windows, turn it off first.

How does Zorlo relate to llama.cpp?

Zorlo uses the ggml library from llama.cpp (MIT license) for computation and produces the same output. It adds expert streaming: it determines which experts each token needs, loads exactly those, and keeps frequently used ones in memory.

Who develops the model?

Qwen3-30B-A3B is developed by the Qwen team at Alibaba Cloud and licensed under Apache 2.0. Zorlo reorganizes how the model is stored on disk and, for the optional fast mode, creates a compact 2-bit copy of its experts.

About

Haider Md. Faiyaz Hussain Shahi

Founder & Lead Developer

Zorlo was founded and is developed by Haider Md. Faiyaz Hussain Shahi. Large AI models have usually required expensive hardware, while most people use an ordinary PC. Zorlo is designed and tested for that PC: it runs a 30-billion-parameter model in 8 GB of memory, with output identical to the full model. Development continues on support for more models and further speed improvements.

Questions, feedback, business
[email protected]
Help with Zorlo
[email protected]