KEYO Studio is a local LLM/SLM runner for developers. Inspect and own model execution on your machine, powered by our JavaScript CPU inference engine—not a wrapper around Ollama, llama.cpp, or vLLM.
Download size: about 158 MB for Windows. Both links deliver the same portable ZIP. If GitHub is slow, try the Hugging Face mirror; speed depends on your connection and location. The first test model is a separate download of about 1 GB. No models are bundled.
Slow or interrupted download? Use your browser’s Downloads menu to pause or resume when supported. If you switch hosts, start a fresh file—do not combine partial downloads. Extract All, keep the files together, then open KEYO Studio.exe. No separate Node.js installation is needed for the Windows app.
Developer alpha 0.1.0-alpha.5. Ugandan-built by KEYO Technologies. Unsigned Windows portable and Linux desktop archives are available with our own CPU engine and model-download library. Extract the entire archive before opening the app. Windows runtime execution is not certified by the Linux build host. CLI requires Node.js 20+; desktop packages include their runtime.
Explore the workspace interface preview →Ugandan-built by KEYO Technologies, KEYO Studio aims to contribute to the AI revolution and add value to the development of AI and future superintelligence.
Desktop builds use our own local CPU engine. Models download separately. None of these archives installs an iPhone or iPad app.
| Platform | Download and instructions | Verification status |
|---|---|---|
| Windows x64 | Portable ZIP, about 158 MB. Windows instructions. | Unsigned developer build; execution on Windows is unverified. |
| Linux x64 desktop | Archive, about 123 MB. Linux instructions. Requires a graphical desktop—not ARM or a headless VPS. | Packaged local model-download/chat smoke passed; not every distribution is certified. |
| macOS Apple Silicon / Intel | Separate arm64 / x64 previews, about 130 / 134 MB. Mac instructions. No separate Node.js installation. | Must finish signing on a Mac. Not Developer ID signed or notarized; Mac GUI execution is unverified. |
| iOS iPhone / iPad | No mobile application yet. | Desktop archives do not run on iOS. Safari offers only an interface preview—not local inference. |
Download the Linux desktop archive — about 123 MB. Open Terminal and run these commands as your normal user—not with sudo.
cd ~/Downloads tar -xzf KEYO-Studio-0.1.0-alpha.5-linux-x64.tar.gz cd KEYO-Studio-0.1.0-alpha.5-linux-x64 ./keyo-studio
Keep the extracted files together. In Model library, select Qwen2.5 0.5B Instruct → Download model → accept the licence and bandwidth confirmation → wait for verification → Load downloaded model → create a conversation. Try “2+2=? Reply with just the number.” The model downloads separately (about 1 GB); afterwards, generation works offline without an account or AI API credits.
Missing library or sandbox error? Report the exact message and Linux distribution/version. Do not disable the sandbox or run as root to bypass an error.
macOS 13 or newer. Apple menu → About This Mac: an Apple M-series chip needs arm64 (about 130 MB); an Intel processor needs x64 (about 134 MB). Use the matching archive, not the Windows or Linux package.
Extract the entire ZIP and read READ-ME-FIRST.txt for the runtime’s minimum macOS version. These cross-built previews require a local signing step: open Terminal, change into the extracted KEYO-Studio folder, then run bash finish-on-mac.command. The script only applies local ad-hoc code signatures; it is not Apple notarization and does not bypass Gatekeeper. After it succeeds, open KEYO Studio.app and follow the same Model library → download → load → new conversation steps above.
Not a certified Mac release. If macOS blocks opening or the signing step fails, report the exact message and Mac processor/macOS version. Do not disable Gatekeeper, remove quarantine, or bypass security protections. A real Mac GUI test and Apple-signed/notarized distribution remain separate requirements.
Portable means there is no installer wizard. Use Windows x64. This is an unsigned developer alpha; execution on Windows is not yet certified.
Use either Windows app button above. The filename must be KEYO-Studio-0.1.0-alpha.5-win32-x64.zip (about 158 MB). Source code, source.tgz, github-upload and release-files ZIPs are for developers—not the runnable Windows app. Wait for the download to finish.
In Downloads, right-click the ZIP, select Extract All, then Extract. Open the extracted folder and double-click KEYO Studio.exe. Do not run it inside the ZIP or move only the EXE. Keep the other files beside it. No separate Node.js installation is needed.
Select Qwen2.5 0.5B Instruct, click Download model, and review the licence and bandwidth confirmation. Keep Internet connected until downloading and checksum verification finish. Allow about 1 GB for this separate model download. Start with 0.5B; 7B and 14B operation is not certified.
Click Load downloaded model and wait for the loaded status. Create a new conversation, then send: “2+2=? Reply with just the number.” The expected answer is 4. CPU generation can be slow; allow it to finish or use the app’s cancellation control.
After downloading and loading, disconnect Wi-Fi and ask another short question. Generation runs locally without an AI API key or inference credits. Close and reopen the app to check saved conversations; load the downloaded model again if needed. No account is required for these local tests.
No EXE? Check the ZIP filename in step 1. Slow download? Try the alternate host or resume through your browser when supported. Windows warning, missing DLL or launch failure? Keep the exact message and report it with your Windows version. Do not disable antivirus or other security protections. Compare SHA-256 with the release checksum file if you need to verify the archive.
KEYO loads supported model assets from a folder you provide, runs our transformer engine on your CPU, and exposes a versioned API bound to localhost with authentication. It is a small, inspectable starting point for local AI development.
Prompts are processed locally. Explicit model downloads contact Hugging Face only after you accept the licence and bandwidth use. Downloads are pinned and checksum-verified; inference works offline afterwards. No paid inference API or telemetry is required.
The engine implements basic GPT-Neo, Qwen2 and Llama architectures. One small public Llama checkpoint has produced text locally; other families have mathematical fixture tests, not pretrained certification. Your existing Afro AI model is not yet certified.
Desktop: extract the archive, open Model library, download Qwen2.5 0.5B, then load it and create a conversation. CLI: use keyo models and keyo download with explicit licence acceptance. Models are not included in the app download.
# Node.js 20 or newer
$ npm install -g ./afro-ai-keyo-studio-0.1.0-alpha.5.tgz
$ keyo inspect /path/to/model
$ keyo chat /path/to/model --prompt "Hello"
$ keyo serve /path/to/modelThe alpha expects these files to already exist together in a supported local model directory. Missing files or incompatible model configs can prevent a model from loading.
A compatible model directory with a single safetensors checkpoint up to 4 GiB, or experimental indexed shards up to 32 GiB total (8 GiB per shard). KEYO reads matrix blocks from disk instead of decoding all weights into RAM. Its optional KEYO-owned native CPU kernel requires a local source build. Context stays limited to 512 tokens. GPU, GGUF and actual 7B/14B checkpoint certification remain unfinished. KEYO does not bundle private weights, automatically fetch them, or send prompts to a cloud service.
Required files · local only · not included in the package
We’d rather name the edges clearly than imply broad compatibility. The alpha is a focused local execution experiment for developers.
Basic GPT-Neo, Qwen2 and Llama execution using our JavaScript CPU engine, with explicitly limited variants and model sizes.
GPU acceleration is roadmap work, not a capability of this release.
Safetensors weights and BPE/tokenizer assets from a provided model folder.
Quantization, sharded checkpoints and large-model certification remain future work.
A versioned API intended for local development and inspection.
Training and fine-tuning are not part of this alpha.
Execution stays local; there is no cloud prompt service or telemetry collection.
Windows portable ZIP and Linux archive include our own engine and model library. Windows/macOS execution certification, signing, GPU acceleration and general 14B support remain unfinished. Never disable security protections to run a package.
To run on Windows, use a Windows app link above. Developer source and release-file bundles below are not Windows installers; they require Node.js 20 or newer.
Early developer alpha · Source package · Model weights not included
Published developer source: GitHub — KEYO Studio. SLM/LLM roadmap. Live browser interface preview: Hugging Face — KEYO Studio. The preview does not run models.