# KEYO Studio v0.1.0-alpha.5

**Developer alpha · Unsigned desktop packages and source · MIT code licence**

Ugandan-built by KEYO Technologies. A local LLM/SLM runner with our own
JavaScript CPU transformer engine, not an Ollama, llama.cpp or vLLM wrapper.

## Included
- CLI for model inspection, local generation and an authenticated localhost API.
- Electron desktop workspace source with local conversation history.
- Persistent System/Light/Dark appearance and Automatic/manual language settings
  for all 15 supported languages, including Arabic right-to-left layout.
- Device-language detection works offline; explicit choices take priority.
- Translated controls, model-library guidance and download/delete confirmations.
- Pinned public model catalogue with licence consent, verified downloads,
  progress, cancellation/resume, private local storage and confirmed deletion.
- Desktop download → load → local chat flow. Models are not bundled.
- Windows x64 portable ZIP and Linux x64 archive attached separately when built.
  Unsigned developer builds: see package metadata for execution-verification status.
- Apple Silicon and Intel macOS preview ZIPs require local signing on a Mac.
  They are not Developer ID signed, notarized or GUI-certified. iOS is not supported.
- Basic GPT-Neo, Qwen2 and Llama execution for explicitly supported variants.
- Safetensors F32/F16/BF16 loading and byte-level BPE tokenization.
- Experimental indexed shards with strict path, tensor and size checks.
- Optional KEYO-owned C++ CPU block kernel, explicitly selected; source only.
- Local instruction-smoke/benchmark tool with optional JavaScript/native parity.
- CI checks the source suite on Linux, Windows and macOS with Node.js 22 and 24.
  Check the run for this release; CI is not signed-installer certification.

## Requirements and limits
- CLI: Node.js >=20. Desktop development: Node.js >=22.12 and Electron.
- Supply a compatible model folder containing config.json, tokenizer.json and
  model.safetensors or model.safetensors.index.json and its shards, or explicitly
  download a pinned model through the library. No model weights are included.
- Single-file checkpoints up to 4 GiB; above 256 MiB, matrices stay on disk and
  are read through a bounded 256 KiB buffer. Resident vectors are capped at 8 MiB.
- Indexed checkpoints: 8 GiB per shard, 32 GiB total, at most 128 shards.
  Each shard has a 256 KiB scratch buffer; the 8 MiB vector cap is shared.
- Native source builds: Linux/macOS with C++ and Node development headers.
- First desktop test: extract the entire archive, open KEYO Studio, expand
  Model library, select Qwen2.5 0.5B, accept the licence/download, then Load downloaded model.
  Create a conversation and ask a short question. Prompts stay local.
  Windows uses JavaScript. Native operation is opt-in, without silent fallback.
- NFC byte-level BPE, exact literal added tokens and modern default RoPE settings.
- Existing Afro AI checkpoint support is a CPU smoke check, not certification
  of quality, speed, all Qwen variants or 7B/14B support.
- Context: 512 tokens. Maximum generated output: 128 tokens.
- A small public Llama checkpoint and the existing merged Afro AI Qwen2
  checkpoint have been smoke-tested. This does not establish numerical parity,
  useful chat quality, production performance or 7B/14B support.

## Not included
GPU inference, GGUF/integer quantization, certified 7B/14B execution, training,
unattended model downloads, cloud inference or signed desktop installers.
There is no prompt telemetry or paid inference API.

## Download files
- afro-ai-keyo-studio-0.1.0-alpha.5.tgz: npm-installable source package.
- KEYO-Studio-0.1.0-alpha.5-source.zip: developer source, licence, tests and CI workflow.
- KEYO-Studio-0.1.0-alpha.5-SHA256SUMS.txt: SHA-256 checksums for these source files and these notes.

Install the CLI with:

    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"

The code licence does not grant rights to third-party model weights.

## Public links
- GitHub: https://github.com/kayondoabass-svg/keyo-studio
- Hugging Face: https://huggingface.co/spaces/kayondoabass/KEYO-Studio

The Hugging Face Space is a browser-local interface preview and source
download. Model loading and generation are deliberately disabled there.