File size: 8,698 Bytes
bc24c5a bad5cb1 bc24c5a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 | ---
base_model:
- Qwen/Qwen2.5-Coder-32B-Instruct
language:
- en
license: mit
pipeline_tag: text-generation
tags:
- code
- coding-agent
- SWE-agent
- distillation
- agent
library_name: transformers
---
<h1 style="
font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Helvetica,Arial,sans-serif;
font-size:48px;
font-weight:700;
line-height:1.25;
text-align:center;
margin:0 0 24px;">
Mocha-Coder-32B
</h1>
<p style="text-align:center; margin:0 0 8px; font-size:16px;">
<a href="https://junliwang.tech/">Junli Wang</a><sup>*</sup>
<a href="https://blankcheng.github.io/">Zhoujun Cheng</a><sup>*†</sup>
<a href="https://yuxuan-zhang-dexter.github.io/">Yuxuan Zhang</a><sup>*</sup>
<a href="https://ber666.github.io/">Shibo Hao</a>
<a href="https://yaotang23.github.io/">Yao Tang</a>
<br>
<a href="https://zhiting.ucsd.edu/">Zhiting Hu</a>
<a href="https://prithvirajva.com/">Prithviraj Ammanabrolu</a>
<a href="https://haozhang.ai/">Hao Zhang</a><sup>†</sup>
</p>
<p style="text-align:center; margin:0 0 24px; font-size:14px; color:#555;">
University of California, San Diego ·
<sup>*</sup>Equal Contribution ·
<sup>†</sup>Corresponding Author
</p>
<div style="
display:flex;
justify-content:center;
gap:12px;
flex-wrap:wrap;
margin-bottom:28px;">
<a href="https://github.com/cocoa-org/NanoRollout" style="
display:inline-block;
padding:8px 24px;
background:#2b2b2b;
color:#ffffff;
border-radius:36px;
text-decoration:none;
font-weight:600;
font-size:16px;">
🧑💻 NanoRollout Code
</a>
<a href="https://huggingface.co/ZeonLap/Mocha-Coder-32B" style="
display:inline-block;
padding:8px 24px;
background:#2b2b2b;
color:#ffffff;
border-radius:36px;
text-decoration:none;
font-weight:600;
font-size:16px;">
🤗 Mocha-Coder-32B Model
</a>
<a href="https://cocoa-org.notion.site/nanorollout" style="
display:inline-block;
padding:8px 24px;
background:#2b2b2b;
color:#ffffff;
border-radius:36px;
text-decoration:none;
font-weight:600;
font-size:16px;">
📒 Blog
</a>
</div>
<div style="max-width:900px;margin:0 auto;">
# Introduction
<div style="
max-width: 880px;
margin: 0 auto;
text-align: justify;
text-justify: inter-word;
line-height: 1.6;">
**Mocha-Coder-32B** is a strong open-data coding agent built on top of **Qwen2.5-Coder-32B-Instruct**. It is trained entirely through distillation on a 300K+ trajectory mixture sampled with our lightweight agent-rollout infrastructure, **NanoRollout**, with no reinforcement learning. The full training signal comes from frontier open-source teacher models (Qwen3-Coder-480B-A35B, Kimi-K2.5, Qwen3-Coder-Next, DeepSeek-V3.2) generating trajectories across multiple agent harnesses (OpenHands, mini-swe-agent, Terminus-2 JSON) on SWE-Rebench, SWE-Smith, and SETA.
The result is a simple but strong baseline coding agent: at the ≤32B scale, Mocha-Coder-32B is the state-of-the-art among open-data models and is competitive with much larger open-source models on agentic SWE benchmarks.
</div>
### Key Features
- **Strong agentic SWE performance**: 62.6 Pass@1 on SWE-Bench Verified, 35.3 on SWE-Bench Pro, 23.6 on Terminal-Bench 2.0, competitive with Qwen3-Coder-480B-A35B-Instruct.
- **Multi-harness training**: Trajectories cover OpenHands, mini-swe-agent, and Terminus-2 JSON, mitigating harness-specific overfitting.
- **Open data**: Distilled from a fully released 300K+ trajectory mixture (`ZeonLap/Mocha-trajectories`).
# Performance
### SWE-Bench Verified
<div align="center">
| **Model** | **Max Iteration** | **SWE-Bench Verified (Pass@1)** |
|----------------------------------|:-----------------:|:-------------------------------:|
| Qwen3-Coder-480B-A35B-Instruct | 100 | 67.0 |
| **Mocha-Coder-32B** | 100 | **62.6** |
| SWE-Master-32B-RL | 150 | 61.4 |
| Kimi-Dev-72B | Agentless, TTS@40 | 60.4 |
| CoderForge-Preview-32B | 100 | 59.4 |
| GLM-4.7-Flash | 100 | 59.2 |
| daVinci-Dev-72B | 100 | 58.5 |
| daVinci-Dev-32B | 100 | 56.1 |
| SERA-32B | 100 | 54.2 |
| Qwen3-Coder-30B-A3B-Instruct | 100 | 51.6 |
| Qwen2.5-Coder-32B-Instruct (Base)| 100 | 6.2 |
</div>
### SWE-Bench Pro
<div align="center">
| **Model** | **Max Iteration** | **SWE-Bench Pro (Pass@1)** |
|----------------------------------|:-----------------:|:--------------------------:|
| Qwen3-Coder-480B-A35B-Instruct | 250 | 38.7 |
| **Mocha-Coder-32B** | 250 | **35.3** |
| Gemini-3-flash | 250 | 34.6 |
| Kimi-K2-Instruct | 250 | 27.7 |
| DeepSeek-V3.2 | 250 | 15.6 |
| Qwen2.5-Coder-32B-Instruct (Base)| 250 | 0.0 |
</div>
### Terminal-Bench 2.0
<div align="center">
| **Model** | **Terminal-Bench 2.0** |
|----------------------------------|:----------------------:|
| Qwen3-Coder-480B-A35B-Instruct | 23.9 |
| **Mocha-Coder-32B** | **23.6** |
| Qwen3-Coder-30B-A3B-Instruct | 13.5 |
| Qwen2.5-Coder-32B-Instruct (Base)| 3.4 |
</div>
# Training Data
Mocha-Coder-32B is trained on a **300K+ trajectory** distillation mixture, drawn from previously released distillation sets (120K) and trajectories newly generated with NanoRollout (~180K).
| **Dataset** | **Teacher Model** | **Harness** | **# Trajectories (K)** | **Source** |
|-----------------|-----------------------------|-------------------|:----------------------:|-------------------|
| SWE-Rebench | Qwen3-Coder-480B-A35B | OpenHands | 32.2 | Nebius |
| SWE-Smith | Qwen3-Coder-480B-A35B | OpenHands | 89.5 | CoderForge |
| SWE-Rebench | Kimi-K2.5 | mini-swe-agent | 83.6 | NanoRollout |
| SWE-Rebench | Qwen3-Coder-Next | mini-swe-agent | 11.5 | NanoRollout |
| SWE-Smith | Qwen3-Coder-480B-A35B | mini-swe-agent | 12.8 | NanoRollout |
| SWE-Smith | Qwen3-Coder-Next | mini-swe-agent | 9.1 | NanoRollout |
| SETA | Kimi-K2.5 / DeepSeek-V3.2 | Terminus-2 JSON | 14.0 | NanoRollout |
The full mixture is released at [`ZeonLap/Mocha-trajectories`](https://huggingface.co/datasets/ZeonLap/Mocha-trajectories).
# Running as an Agent
Mocha-Coder-32B is trained as an agent and is most useful when paired with a coding-agent harness. We have validated it with:
- **mini-swe-agent** — minimal SWE agent loop, recommended for SWE-Bench Verified / Pro evaluation.
- **OpenHands** — full-featured SWE harness; the model was trained on OpenHands trajectories.
- **Terminus-2 JSON** — for Terminal-Bench 2.0 style shell tasks.
Point each harness's model endpoint at the vLLM server above. For SWE-Bench Verified we report numbers at a 100-iteration budget; for SWE-Bench Pro at 250 iterations.
# License
Mocha-Coder-32B (model weights, training trajectories, and code) is released under the **MIT License** (see `LICENSE`) for research, educational, and commercial use.
# Citation
If you use Mocha-Coder-32B or NanoRollout in your research, please cite NanoRollout:
```bibtex
@misc{nanorollout,
title = {NanoRollout: A Lightweight Infra for Digital Agent Rollout at Scale},
author = {Wang, Junli and Cheng, Zhoujun and Zhang, Yuxuan and Hao, Shibo
and Tang, Yao and Hu, Zhiting and Ammanabrolu, Prithviraj
and Zhang, Hao},
year = {2026},
howpublished = {\url{https://github.com/cocoa-org/NanoRollout}},
}
```
</div>
|