README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.8-27B
base_model_relation: finetune
datasets:
  - Roy-DOU/Webdev-Bench
tags:
  - agent
  - code
  - frontend
  - web-development
  - reinforcement-learning
---

# WebCoder-27B

[Project page](https://douzj-llm.github.io/WebCoder/) ·
[Code](https://github.com/douzj-llm/WebCoder) ·
[Webdev-Bench](https://huggingface.co/datasets/Roy-DOU/Webdev-Bench) ·
[Paper](https://arxiv.org/abs/2610.23104)

WebCoder-27B is a general frontend development agent, trained from
[Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) with artifact-grounded agentic
reinforcement learning. It is the model released with the paper

**WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic
Reinforcement Learning**
Zijun Dou\*, Zhen Yang\*, Xijun Liu, Dongzhuoran Zhou, Aohan Zeng, Jie Tang
*\*Core contributors.*

## Overview

Real-world frontend development requires an agent to compose reproduction, generation,
modification and repair over an evolving application state. WebCoder assesses all of these
activities through their effects on a shared executable artifact and on its satisfaction of
the task requirements.

- **Webdev-World** — an executable environment that integrates code editing, application
  execution, browser inspection and on-demand skills, preserving project state across
  interactions.
- **Artifact-grounded reward and credit assignment** — an independent verifier evaluates the
  final application against the task requirements to give a terminal reward; changes in
  intermediate artifact progress redistribute the resulting rollout advantage across
  development segments.
- **Long-horizon training** — context compaction and message-tree sample construction support
  training on trajectories beyond a single context window, without repeated loss computation
  on shared generated history.

## Results

| Model | Vision2Web Avg | Webdev-Bench RS | Webdev-Bench Avg | Webdev-Bench DSR |
|---|---:|---:|---:|---:|
| Claude-Opus-4.8 | 62.5 | 65.1 | 73.7 | 100.0% |
| GPT-5.4 | 57.9 | 59.6 | 70.8 | 100.0% |
| Qwen3.8-27B (base) | 41.1 | 64.8 | 72.8 | 95.0% |
| **WebCoder-27B** | **62.4** | **72.8** | **79.8** | **100.0%** |

RS is requirement satisfaction and DSR the delivery success rate. The full table, with every
baseline and every sub-metric, is in the paper and on the
[project page](https://douzj-llm.github.io/WebCoder/#results).

## Usage

WebCoder-27B keeps the architecture (`Qwen3_5ForConditionalGeneration`), tokenizer, chat
template and processor configuration of its base model, so it is served the same way as
Qwen3.8-27B — for example with vLLM or SGLang. See the
[Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B) for serving commands and
API examples.

```bash
vllm serve Roy-DOU/WebCoder-27B --tensor-parallel-size 4
```

`generation_config.json` sets `temperature=1.0`, `top_p=0.95`, `top_k=20`.

The model was trained as an agent: it acts through tools in a development environment over
many turns, and is at its best in that setting rather than in single-turn chat.

## License

Released under the Apache License 2.0, the license of the base model. See `LICENSE`.

## Citation

```bibtex
@article{dou2026webcoder,
  title   = {WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning},
  author  = {Dou, Zijun and Yang, Zhen and Liu, Xijun and Zhou, Dongzhuoran and Zeng, Aohan and Tang, Jie},
  journal = {arXiv preprint arXiv:2610.23104},
  year    = {2026}
}
```