Zen Coder: Code Generation That Actually Works

Most code generation models are trained on code. Zen Coder is trained on software engineering - the entire workflow from requirements to deployed software.

Beyond Code Completion

Traditional code models complete syntax:

# Traditional model
def calculate_total(items):
    # Model: "return sum(items)"

Zen Coder understands context:

# Zen Coder
# Given: e-commerce codebase, tax calculation requirements, existing patterns
def calculate_total(items, shipping, customer):
    subtotal = sum(item.price * item.quantity for item in items)
    tax = calculate_tax(subtotal, customer.shipping_address)
    shipping_cost = get_shipping_cost(items, shipping.method)
    discount = apply_promotions(items, customer.loyalty_tier)
    return OrderTotal(
        subtotal=subtotal,
        tax=tax,
        shipping=shipping_cost,
        discount=discount,
        total=subtotal + tax + shipping_cost - discount
    )

The model understands the domain, not just the syntax.

Training Data

We trained on:

Commit History: Not just final code, but the evolution of code over time.

Pull Requests: Complete context including discussion, review, iteration.

Issue → Code: How requirements become implementations.

Test → Code: How tests inform implementation.

Documentation: How engineers explain what code does.

This gives the model understanding of why code is written, not just what it looks like.

Model Architecture

Zen Coder is a family of models:

Model	Parameters	Context	Specialization
Zen Coder Nano	600M	4K	Fast completion
Zen Coder	7B	32K	General purpose
Zen Coder Pro	34B	128K	Complex reasoning
Zen Coder Max	72B	256K	Architecture-level

All models share the same training methodology, scaled to different capabilities.

Key Capabilities

Multi-File Understanding

User: Add authentication to the API

Zen Coder:
- Reads existing route structure
- Identifies authentication patterns in codebase
- Modifies routes.py, middleware.py, config.py
- Updates tests
- Adds migration for user tokens

Test-Driven Generation

User: Write a function that passes these tests
[test_cases.py content]

Zen Coder:
- Analyzes test cases
- Infers function signature and behavior
- Generates implementation
- Runs tests to verify

Codebase-Aware Refactoring

User: Refactor the payment module to use the new API

Zen Coder:
- Maps all usages of old API
- Generates refactored code for each
- Preserves behavior through tests
- Handles edge cases

IDE Integration

Zen Coder integrates with your development environment:

VS Code Extension: Inline suggestions, chat, multi-file edits.

JetBrains Plugin: Full IDE integration for IntelliJ, PyCharm, etc.

Vim/Neovim: hanzo.vim plugin for terminal-native development.

CLI: Command-line interface for scripting and automation.

# Ask Zen Coder from terminal
zen-coder "Add rate limiting to all API endpoints"

Benchmarks

Benchmark	GPT-4	Claude 3	Zen Coder 34B
HumanEval	87.1%	84.9%	89.3%
MBPP	83.5%	85.2%	88.1%
SWE-bench	33.2%	37.8%	42.6%

SWE-bench tests real software engineering tasks - fixing actual GitHub issues. This is where training on workflows matters.

What Makes It Different

1. Context Length: 256K tokens means entire codebases in context
2. Multi-file: Understands relationships across files
3. Iteration: Learns from feedback, improves over conversation
4. Testing: Generates and runs tests automatically
5. Style Matching: Follows existing code style in the project

The Philosophy

Code generation isn't about typing faster. It's about amplifying engineering judgment.

Zen Coder handles the routine parts - boilerplate, standard patterns, mechanical changes - so engineers can focus on the interesting problems: architecture, algorithms, user experience.

The best code generator is one that makes you a better engineer, not one that replaces engineering.

---

This post is part of our retrospective series exploring the technical foundations of Hanzo and Zen.