zen-docs/content/docs/models/zen-coder.mdx

---
title: Zen Coder
description: Agentic coding AI from edge to frontier — 4B, 31B MoE, and 480B MoE variants
---

# Zen Coder

Zen Coder is Zen LM's family of code-focused AI models spanning three capability tiers. The flagship `zen-coder-480b-instruct` is a 480B-parameter Mixture of Experts model with 35B active parameters, delivering state-of-the-art results on agentic coding, browser-use, and tool-use benchmarks. The family supports 358 programming languages, fill-in-the-middle (FIM), function calling, and extended chain-of-thought reasoning.

## Model Family

| Model | Parameters | Active | Context | Use Case |
|-------|------------|--------|---------|----------|
| [zen-coder](https://huggingface.co/zenlm/zen-coder) | 4B | 4B | 32K | Edge / mobile |
| [zen-coder-flash](https://huggingface.co/zenlm/zen-coder-flash) | 31B MoE | 3B | 131K | Balanced performance |
| [zen-coder-480b-instruct](https://huggingface.co/zenlm/zen-coder-480b-instruct) | 480B MoE | 35B | 256K | Frontier agentic coding |

## Highlights

- State-of-the-art on SWE-bench agentic coding benchmarks
- 256K context (extendable to 1M via Yarn) — handles full repository scale
- 358 programming languages (ABAP to Zig)
- Fill-in-the-middle (FIM) code insertion at cursor position
- Function calling with structured tool-use
- Extended chain-of-thought reasoning for complex engineering tasks

## Quick Start

### Flagship (480B)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "zenlm/zen-coder-480b-instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-coder-480b-instruct")

messages = [{"role": "user", "content": "Write a quicksort algorithm in Python with type hints."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=4096)
output = tokenizer.batch_decode(
    [generated_ids[0][len(inputs.input_ids[0]):]],
    skip_special_tokens=True
)[0]
print(output)
```

### Edge (4B)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-coder", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-coder")

messages = [{"role": "user", "content": "Write a binary search in Rust."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(output[0][len(inputs.input_ids[0]):], skip_special_tokens=True))
```

## Fill-in-the-Middle (FIM)

```python
input_text = """<|fim_prefix|>def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[len(arr) // 2]
    <|fim_suffix|>
    middle = [x for x in arr if x == pivot]
    right = [x for x in arr if x > pivot]
    return quicksort(left) + middle + quicksort(right)<|fim_middle|>"""

messages = [
    {"role": "system", "content": "You are a code completion assistant."},
    {"role": "user", "content": input_text}
]
```

## Use Cases

- Agentic coding and software engineering automation
- Full-repository context understanding
- Code review, debugging, and refactoring
- Multi-language code translation
- Fill-in-the-middle completion in IDEs

## Formats

- **SafeTensors** (BF16) — native precision
- **GGUF** — Q4_K_M, Q5_K_M, Q8_0 for llama.cpp
- **MLX** — Apple Silicon optimized

## Performance Benchmarks

| Benchmark | zen-coder-480b | zen-coder-flash |
|-----------|---------------|-----------------|
| SWE-bench Verified | State-of-the-art | 59.2% |
| AIME 2025 | Competitive | 91.6% |
| GPQA | Competitive | 75.2% |
| t²-Bench | Competitive | 79.5% |

## Hardware Requirements

| Model | VRAM |
|-------|------|
| zen-coder (4B) | 8 GB |
| zen-coder-flash (31B MoE) | 24 GB |
| zen-coder-480b-instruct | 8x 80 GB |

## Get Started

- [HuggingFace](https://huggingface.co/zenlm/zen-coder-480b-instruct)
- [GitHub](https://github.com/zenlm/zen-coder)