# Introduction
**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.
### 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
| **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 |
### SWE-Bench Pro
| **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 |
### Terminal-Bench 2.0
| **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 |
# 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}},
}
```