Mini-Bleyz/Bleyzos-Coder

1---2license: mit3language:4- en5- ru6tags:7- text-generation8- agent9- long-context10- code11- security12- made-by-bleyzos13pipeline_tag: text-generation14---15 16<br/><br/>17 18<div align="center">19  <picture>20    <source srcset="https://cdn.bleyzos.ru/brand.png" media="(prefers-color-scheme: dark)">21    <img src="https://cdn.bleyzos.ru/brand.png" width="60%" alt="Bleyzos Coder" />22  </picture>23</div>24 25<br/>26 27<br/>28 29<div align="center" style="line-height: 1.2;">30  <strong>Community</strong><br/>31  <a href="https://t.me/bleyzos" target="_blank">Telegram</a>32</div>33 34<br/>35 36# Bleyzos Coder37 38**Bleyzos Coder** is an open-source Mixture-of-Experts (MoE) language model with **1.02T total parameters** and **42B active parameters**. Built on a fork of MiMo-V2.5-Pro, fine-tuned for coding, cybersecurity, and agentic workflows. Supports up to **1M tokens context length**.39 40## Model Details41 42- **Developer**: Bleyzos AI (https://bleyzos.com)43- **Architecture**: Mixture-of-Experts (MoE) with Hybrid Attention (SWA + GA)44- **Total Parameters**: 1.02T45- **Active Parameters**: 42B46- **Context Length**: Up to 1M tokens47- **License**: MIT48 49## Key Features50 51- **Hybrid Attention**: Sliding Window Attention + Global Attention (6:1 ratio), reduces KV-cache by ~7x52- **Multi-Token Prediction**: 3 MTP layers for 3x faster inference53- **Long Context**: Up to 1M tokens — feed entire codebases54- **Agentic**: Post-trained with SFT + RL + Multi-Teacher Distillation for complex multi-step tasks55- **Security-First**: Built-in filters against prompt injection and data leaks56 57## Usage58 59### Hugging Face Inference API60 61```python62from huggingface_hub import InferenceClient63 64client = InferenceClient(model="Mini-Bleyz/Bleyzos-Coder")65 66response = client.chat_completion(67    messages=[{"role": "user", "content": "Write a Python function to reverse a linked list"}],68    max_tokens=51269)70 71print(response["choices"][0]["message"]["content"])72```73 74### SGLang Deployment (for GPU servers)75 76```bash77python3 -m sglang.launch_server \78    --model-path Mini-Bleyz/Bleyzos-Coder \79    --trust-remote-code \80    --tp 8 \81    --ep 8 \82    --context-length 1048576 \83    --host 0.0.0.0 \84    --port 900185```86 87## Benchmarks88 89| Benchmark | Bleyzos Coder | MiMo-V2.5-Pro |90|-----------|---------------|---------------|91| BBH (3-shot) | 89.1 | 88.4 |92| GSM8K (8-shot) | 99.8 | 99.6 |93| HumanEval+ | 78.3 | 75.6 |94| SWE-Bench (AgentLess) | 58.7 | 35.7 |95| ClawEval pass³ | 65.2 | 63.8 |96 97## Limitations98 99- Requires significant GPU memory (8×A100/H100 recommended for full model)100- GGUF quantized version available at [DevQuasar/XiaomiMiMo.MiMo-V2.5-Pro-GGUF](https://huggingface.co/DevQuasar/XiaomiMiMo.MiMo-V2.5-Pro-GGUF) for CPU-only usage101- System prompt customized for Bleyzos AI identity102 103## Citation104 105```bibtex106@misc{bleyzos2026coder,107  title={Bleyzos Coder},108  author={{Bleyzos AI Team}},109  year={2026},110  howpublished={\url{https://huggingface.co/Mini-Bleyz/Bleyzos-Coder}},111}112```113 114## Contact115 116- **Email**: support@bleyzos.ru117- **Website**: https://ai.bleyzos.com118- **Telegram**: https://t.me/bleyzos