README.md

---
license: mit
base_model:
- ByteDance-Seed/Seed-Coder-8B-Base
---

# Seed-Coder-8B-Reasoning

## Introduction
**Seed-Coder-8B-Reasoning** is an 8-billion-parameter model further optimized for **code reasoning**, **problem-solving**, and **algorithmic thinking** tasks.  
Built upon the strong base of Seed-Coder, it undergoes additional training in sandbox environments to significantly enhance its ability to tackle complex coding problems and competitions. It features:
- Trained on a **massively curated corpus**, filtered using an **LLM-based method** to ensure high-quality real-world code, text-code alignment, and synthetic datasets.
- **Sandbox fine-tuning** to specifically strengthen **multi-step reasoning**, **algorithm design**, and **competitive programming** capabilities.
- Maintains **long-context handling** up to 32K tokens, enabling it to reason over extended problem descriptions and large input-output examples.

<p align="center">
  <img width="100%" src="imgs/large_model_comparison_split-crop.png">
</p>

## Model Downloads
| Model Name                  | Length | Download   |    Notes |
|---------------------------------------------------------|-----------|------------------------------------|-----------------------|
| Seed-Coder-8B-Base           | 32K    | 🤗 [Model](https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base)   |  Pretrained on our model-centric code data.  |
| Seed-Coder-8B-Instruct             | 32K    | 🤗 [Model](https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Instruct)   |  Instruction-tuned for alignment with user intent. |
| 👉 **Seed-Coder-8B-Reasoning**            | 32K    | 🤗 [Model](https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Reasoning)   |  RL trained to boost reasoning capabilities.  |


## Requirements
You will need to install the latest versions of `transformers` and `accelerate`:

```bash
pip install -U transformers accelerate
```

## Quickstart

Here is a simple example demonstrating how to load the model and perform code generation using the Hugging Face `pipeline` API:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "ByteDance-Seed/Seed-Coder-8B-Reasoning"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)

messages = [
    {"role": "user", "content": "Write a quick sort algorithm."},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_tensors="pt",
    add_generation_prompt=True,  
).to(model.device)

outputs = model.generate(input_ids, max_new_tokens=16384)
response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)
```

## Evaluation

Seed-Coder-8B-Reasoning has been evaluated extensively on reasoning-intensive code benchmarks, showing:
- Significant improvements on **competitive programming** datasets and coding challenges.
- Enhanced ability to **break down complex problems**, **design correct algorithms**, and **produce efficient implementations**.
- Strong generalization to unseen problems across multiple domains (math, strings, arrays, graphs, DP, etc.).

<table>
  <tr>
    <th rowspan="2">Model</th>
    <th colspan="3">LiveCodeBench-Hard</th>
    <th colspan="3">LiveCodeBench-Medium</th>
    <th colspan="3">LiveCodeBench-Easy</th>
    <th rowspan="2">Overall</th>
  </tr>
  <tr>
    <th>4mon</th><th>3mon</th><th>2mon</th>
    <th>4mon</th><th>3mon</th><th>2mon</th>
    <th>4mon</th><th>3mon</th><th>2mon</th>
  </tr>

  <!-- ~8B Models -->
  <tr><td colspan="11"><b>~8B Models</b></td></tr>
  <tr>
    <td>DeepSeek-R1-Distill-Qwen-7B</td>
    <td>11.3</td><td>10.7</td><td>9.6</td>
    <td>39.6</td><td>37.2</td><td>37.1</td>
    <td>76.2</td><td>77.1</td><td>67.1</td>
    <td>36.5</td>
  </tr>
  <tr>
    <td>DeepSeek-R1-Distill-Seed-Coder-8B</td>
    <td>13.6</td><td>13.9</td><td>13.4</td>
    <td>39.6</td><td>38.7</td><td>39.3</td>
    <td>79.8</td><td>80.2</td><td>73.2</td>
    <td>39.0</td>
  </tr>
  <tr>
    <td>OlympicCoder-7B</td>
    <td>12.7</td><td>11.8</td><td>12.5</td>
    <td>40.8</td><td>39.0</td><td>38.7</td>
    <td>78.0</td><td>77.1</td><td>67.8</td>
    <td>37.9</td>
  </tr>
  <tr>
    <td>Qwen3-8B-thinking</td>
    <td>27.5</td><td>23.5</td><td>19.7</td>
    <td>65.7</td><td>59.7</td><td>58.5</td>
    <td>98.0</td><td>98.1</td><td>97.3</td>
    <td>57.4</td>
  </tr>
  <tr>
    <td>Seed-Coder-8B-Reasoning</td>
    <td>27.6</td><td>28.0</td><td>31.0</td>
    <td>65.8</td><td>59.2</td><td>57.5</td>
    <td>87.8</td><td>88.0</td><td>80.1</td>
    <td>53.6</td>
  </tr>

  <!-- 13B+ Models -->
  <tr><td colspan="11"><b>13B+ Models</b></td></tr>
  <tr>
    <td>DeepSeek-R1-Distill-Qwen-14B</td>
    <td>21.3</td><td>20.5</td><td>16.1</td>
    <td>58.1</td><td>53.4</td><td>51.4</td>
    <td>93.3</td><td>94.2</td><td>93.7</td>
    <td>51.9</td>
  </tr>
  <tr>
    <td>Claude-3.7-Sonnet-thinking</td>
    <td>27.3</td><td>30.8</td><td>31.0</td>
    <td>54.5</td><td>55.1</td><td>51.4</td>
    <td>96.2</td><td>100.0</td><td>100.0</td>
    <td>53.3</td>
  </tr>
  <tr>
    <td>o3-mini-low</td>
    <td>30.3</td><td>32.3</td><td>28.6</td>
    <td>69.6</td><td>61.2</td><td>54.1</td>
    <td>98.7</td><td>100.0</td><td>100.0</td>
    <td>59.4</td>
  </tr>
</table>


For detailed benchmark performance, please refer to our [📑 technical report](https://github.com/ByteDance-Seed/Seed-Coder/blob/master/Seed-Coder.pdf).