richardyoung/
deepseek-coder-33b-heretic:latest

167 3 weeks ago

State-of-the-art coding AI trained on 2T tokens with project-level understanding and no content restrictions for unrestricted code generation.

3 weeks ago

9483261a48e3 · 67GB ·

model
llama
·
33.3B
·
F16
template
{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user
system
# DeepSeek-Coder-33B-Instruct Heretic - Uncensored Edition ## Overview This is an **uncensored versi
params
{ "stop": [ "<|im_start|>", "<|im_end|>" ], "temperature": 0.7, "top

Readme

DeepSeek Coder 33B Heretic: The Ultimate Coding AI

🚀 Overview

DeepSeek Coder 33B Heretic is a state-of-the-art 33 billion parameter code generation model, specifically designed for unrestricted programming assistance and software development. Built from DeepSeek-Coder-33B-Instruct with Heretic Abliteration applied, this model delivers exceptional coding capabilities while maintaining complete freedom from content restrictions.

🎯 Key Features

  • 33B parameters of specialized coding intelligence
  • 87% code training data with 13% natural language
  • Uncensored outputs - no refusal to generate code for any purpose
  • 16K context window for project-level code understanding
  • State-of-the-art coding performance across multiple languages
  • Fill-in-the-Blank capability for code completion tasks

💻 Programming Capabilities

  • Multi-language Support: Python, JavaScript, Java, C++, Rust, Go, and 100+ languages
  • Code Generation: Complete functions, classes, and full applications
  • Code Review: Advanced debugging and optimization suggestions
  • Algorithm Implementation: Complex data structures and algorithms
  • API Development: REST APIs, GraphQL, and microservices
  • Testing: Unit test generation and test case creation

🏆 Benchmark Performance

DeepSeek-Coder-Base-33B significantly outperforms existing open-source code LLMs: - HumanEval Python: 7.9% better than CodeLlama-34B - HumanEval Multilingual: 9.3% better than CodeLlama-34B - MBPP: 10.8% better than CodeLlama-34B - DS-1000: 5.9% better than CodeLlama-34B

DeepSeek-Coder-Instruct-33B achieves: - HumanEval: Outperforms GPT-3.5-turbo - MBPP: Comparable results with GPT-3.5-turbo

💻 Quick Start

# Basic coding assistance
ollama run richardyoung/deepseek-coder-33b-heretic

# Generate a complete function
ollama run richardyoung/deepseek-coder-33b-heretic "Create a FastAPI endpoint for user authentication with JWT"

# Code review
ollama run richardyoung/deepseek-coder-33b-heretic "Review this code for security vulnerabilities and performance issues: [paste code]"

🛠️ Example Use Cases

Full Application Development

ollama run richardyoung/deepseek-coder-33b-heretic "Build a complete Django web application for e-commerce with user accounts, shopping cart, and payment processing"

Algorithm Implementation

ollama run richardyoung/deepseek-coder-33b-heretic "Implement a red-black tree data structure in Rust with all major operations"

Code Refactoring

ollama run richardyoung/deepseek-coder-33b-heretic "Refactor this messy Python code into clean, readable, and optimized code: [paste code]"

API Development

ollama run richardyoung/deepseek-coder-33b-heretic "Create a GraphQL API schema for a social media platform with user posts and comments"

Database Design

ollama run richardyoung/deepseek-coder-33b-heretic "Design and implement the database schema for a hospital management system"

⚙️ Advanced Configuration

ollama run richardyoung/deepseek-coder-33b-heretic \
  --temperature 0.2 \
  --top-p 0.95 \
  --top-k 50 \
  --repeat-penalty 1.03 \
  "Write production-ready code with proper error handling and documentation"

🔧 Technical Specifications

  • Base Model: DeepSeek-Coder-33B-Instruct
  • Training Data: 2T tokens (87% code, 13% natural language)
  • Context Length: 16K tokens
  • Architecture: Transformer with specialized coding optimizations
  • Quantization: Optimized for efficient inference

📊 Language Support

Tier 1 (Expert Level)

  • Python: 95% accuracy on complex tasks
  • JavaScript/TypeScript: Full-stack development
  • Java: Enterprise applications
  • C++: System programming and algorithms
  • Rust: Memory-safe systems programming

Tier 2 (Advanced)

  • Go: Backend services and microservices
  • C#: .NET applications
  • Swift: iOS development
  • Kotlin: Android development
  • Scala: Functional programming

Tier 3 (Competent)

  • PHP, Ruby, Perl: Web development
  • Shell scripting: Bash, PowerShell
  • SQL: Database queries and optimization
  • Assembly: Low-level programming

💾 System Requirements

Minimum Requirements

  • RAM: 48GB (for Q4 quantization)
  • GPU: RTX 4090 or A100 40GB
  • Storage: 80GB free space

Recommended Setup

  • RAM: 64GB+
  • GPU: A100 80GB or multiple RTX 4090s
  • Storage: 200GB NVMe SSD

🌟 What Makes This Model Special

  1. Specialized Training: 2 trillion tokens of high-quality code data
  2. Project-Level Context: Understands large codebases and dependencies
  3. Uncensored Coding: No restrictions on code generation for any purpose
  4. Fill-in-the-Blank: Excellent at completing partially written code
  5. Real-World Focus: Trained on production code, not just examples

⚠️ Usage Guidelines

This is an uncensored coding model. While it provides unrestricted code generation capabilities, users should: - Ensure code follows security best practices - Review generated code for potential vulnerabilities - Comply with applicable licensing and legal requirements - Implement appropriate testing before production deployment

🏗️ Project Templates

Web Development

ollama run richardyoung/deepseek-coder-33b-heretic "Create a complete React TypeScript application with routing, state management, and testing"

API Development

ollama run richardyoung/deepseek-coder-33b-heretic "Build a RESTful API with authentication, rate limiting, and database integration"

Machine Learning

ollama run richardyoung/deepseek-coder-33b-heretic "Implement a complete ML pipeline for image classification with data preprocessing and model deployment"

🔍 Advanced Features

Code Understanding

  • Large codebase analysis: Understand and modify existing projects
  • Dependency management: Handle complex import structures
  • Code documentation: Generate comprehensive documentation

Security Focus

  • Vulnerability detection: Identify security issues in code
  • Best practices: Enforce coding standards and security guidelines
  • Audit ready: Generate audit-friendly code structures

🤝 Support & Community

  • Base Model: DeepSeek-Coder-33B-Instruct
  • Modifications: Heretic Abliteration v1.0.1
  • Performance: Code generation benchmarks
  • Updates: Continuous improvements and optimizations

📝 License

This model follows the DeepSeek model license. Please refer to the original DeepSeek repository for detailed licensing terms.

🙏 Acknowledgments

  • DeepSeek Team for exceptional code generation capabilities
  • Heretic Community for abliteration technology
  • Open Source Community for training datasets and benchmarks
  • Ollama for making powerful coding AI accessible

Note: This model excels at code generation but requires human review for production deployments. Always test thoroughly before shipping code.

Performance Tip: Use lower temperature (0.1-0.3) for precise code generation; higher values (0.4-0.7) for more creative solutions and explanations.