README / README.md
KVC899's picture
Ki Complete: 7B model, VS Code Marketplace and Open VSX links
f39a609 verified
|
Raw History Blame Contribute Delete
4.56 kB
metadata
title: KvcHub Labs
emoji: πŸ›‘οΈ
colorFrom: blue
colorTo: indigo
pinned: false

KvcHub Labs

On-Device Privacy & High-Speed Neural Intelligence.

🌐 Website β€’ πŸ€— Hugging Face β€’ πŸ“¦ PyPI (kava-privacy) β€’ Ki Complete for VS Code β€’ Ki Complete on Open VSX


πŸ‘‹ About

KvcHub Labs researches and builds ultra-lightweight, hardware-accelerated, zero-trust AI models. Our mission is to make advanced privacy redaction and neural intelligence run 100% locally on-device with zero cloud data leakage and sub-millisecond execution speeds.

🌟 Core Focus Areas

  • πŸ”’ Zero Data Leakage (100% On-Device): Text never leaves local memory, supporting HIPAA, GDPR, and SOC2 compliance.
  • ⚑ Sub-Millisecond Speed: Optimized for ultra-fast local inference without cloud network roundtrips.
  • πŸ§ͺ High-Precision Redaction: Engineered to avoid false positives on scientific constants, subnet masks, legal citations, and financial tickers.
  • 🌍 Zero-Shot Multilingual Detection: Recognizes international names, addresses, and phone numbers out of the box.
  • 🌊 Real-Time Streaming: Sanitizes live streaming LLM outputs token-by-token on the fly.

πŸ€– Models

Model Task Parameters Notes
KvcHub/Ki-Complete-1.5B Code Completion 1.5B The default model of Ki Complete, KvcHub's code autocomplete for VS Code and Antigravity. Fine-tuned from Qwen2.5-Coder-1.5B; GGUF (Q8_0, plus a 4-bit Q4_K_M "Small" for 8 GB computers) and the LoRA. Apache-2.0.
KvcHub/Ki-Complete-7B Code Completion 7B Ki Complete's larger model: the most accurate on real code, about 4 times slower; for 16 GB+ computers. Fine-tuned from Qwen2.5-Coder-7B; one GGUF (Q8_0) plus the LoRA. Apache-2.0.
KvcHub/KaiLu-3-Pro Text Generation 1.0B Newest KaiLu chat model, trained from scratch. One GGUF (Q8_0) for llama.cpp, Ollama and LM Studio, plus safetensors. Apache-2.0.
KvcHub/KaiLu3 Text Generation 0.3B KaiLu 3 base model (329M), trained from scratch.
KvcHub/KaiLu-2-2B-GGUF Text Generation 2B Quantized GGUF build for fast local inference.
KvcHub/KaiLu-2-933M Text Generation 1.0B (933M) Compact generative language model.
KvcHub/Kava-Privacy-2.01M-mlx Token Classification 2.01M MLX-optimized on-device PII redaction model.
KvcHub/Kava-Privacy-636k-mlx Token Classification 636k Ultra-compact MLX model for sub-millisecond PII detection.

Ki Complete

Ki Complete is KvcHub's code autocomplete extension for VS Code and Antigravity: code suggestions as you type, shown in green, from a model that runs on your own computer. Your code never leaves your machine, it works offline, and it's free. It turns off GitHub Copilot's (in VS Code) or Antigravity's own suggestions once, so the two don't compete, with a button to turn them back on.

Install it:

Models, chosen from the Ki status bar menu and downloaded once:

  • KvcHub/Ki-Complete-1.5B (default): quickest, about 0.23 seconds per suggestion on an Apple M5 Pro; also as Ki Complete Small (4-bit, 990 MB)
  • KvcHub/Ki-Complete-7B: most accurate on real code (49.7% / 37.7% of line endings exactly right, against 43.1% / 35.7%), about 1 second per suggestion

More at kvchub.com.