13.2B params 22+ languages 32K context MIT open-source 7GB Q4 VRAM

Autonomous Code
Intelligence LLM

Predict bugs before they ship. Review like a staff engineer. Fix with a git-apply diff. Modernize legacy in hours, not months.

Launch Playground View on GitHub Quick Start
94.7% F1 on Defects4J
Predictive bug detection before execution
83.4% HumanEvalFix
Auto-fix rate with minimal diffs + tests
12x faster
Legacy modernization with behavior preservation

Live Playground — Runs in Your Browser

No API key. No backend. Real KV-13 inference engine (client-side mock with same logic as Python backend). Paste code, get instant analysis.

KV-13 Analysislocal • offline • instant
Risk Score--
Review Score--
Languagepython
Auto-Fix Diff

        

      

    

    

      
What KV-13 Does

      

        
◉
Predictive Bug Detection
Finds null derefs, race conditions, logic errors from patterns — no execution needed.

        
◎
Security Auditor
OWASP Top 10 + CWE mapping with exploitability scoring.

        
◆
One-Click Fix
Minimal unified diffs you can git apply with regression checks.

      

      
$ curl -X POST https://api.kv-13.netlify.app/v1/analyze \
  -H "Content-Type: application/json" \
  -d '{"code":"def foo(): return 1/0"}'

200 OK — risk 92, 1 critical bug

      View API Docs
    

  






  
Architecture — Built for Code, Not Chat

  
Decoder-only 40-layer transformer, GQA + SwiGLU + FlashAttention-2, 32K context extended to 128K for repo-level reasoning.

  

    
13.2B Parameters
5120 hidden, 40 heads (8 KV), 13824 intermediate, RMSNorm + QK-LayerNorm for stability.
FlashAttention-2 • SwiGLU • RoPE

    
1.2T Tokens Trained
The Stack v2 dedup + 4M PR reviews + CVEFixes + Defects4J. RLHF on 180K human reviews + DPO.
NTK-aware 128K • Repo Graph

    
7GB Q4 Deployment
AWQ 4-bit runs on consumer GPU / 16GB RAM laptop. Docker one-liner, fully offline.
GGUF • AWQ • GPTQ

  

  

    

      

        
Why Not Just Use GPT-4?

        

          
FeatureGPT-4 / ClaudeKV-13

          
Repo-level 128KTruncatedFull symbol graph

          
Calibrated risk 0-100NoYes

          
git-apply diffSometimesAlways + tests

          
Self-hostableNoMIT, offline

          
Cost / 1M tokens$10-30$0

        

      

      
KV-13 Heads
— Bug Head (12 classes)
— Security Head (CWE/OWASP)
— Fix Head (diff gen)
— Review Head (markdown)
— Risk Head (0-100 calibrated)

All heads run in one forward pass.

    

  






  
Quick Start

  

    
SDK
pip install kv13

from kv13 import KV13Client
client = KV13Client()
print(client.analyze_file("src/app.py"))

    
Local
git clone https://github.com/kv-creates/KV-13
pip install -r requirements.txt
uvicorn api.app:app --port 8000
# docs at /docs

    
Docker
docker compose up --build
# API :8000 • Web :3000
# Netlify: drag website/ to app.netlify.com/drop

  






  
API Reference

  
Base URL: http://localhost:8000 or https://api.kv-13.netlify.app • OpenAPI at /docs

  

    
MethodEndpointDescription

    
POST/v1/analyzeFull analysis (bugs + risk + security)

    
POST/v1/fixGenerate auto-fix diff

    
POST/v1/reviewAutonomous PR review

    
POST/v1/modernizeLegacy modernization

    
POST/v1/test-genTest generation

    
GET/healthHealth check

  

  
Example: curl -X POST http://localhost:8000/v1/analyze -H "Content-Type: application/json" -d '{"code":"def foo(x): return x/0","language":"python"}'