Step-3.5-Flash-Midtrain-GGUF-by-AesSedai

Step 3.5 Flash Base Midtrain - GGUF

<p align="center" style="font-size: 1.1em; font-weight: bold; margin-top: 15px; margin-bottom: 24px;"> This repository serves as a mirror of the original GGUF created by <a href="https://huggingface.co/AesSedai/models">Aes Sedai</a>. </p> <div align="center"> <div align="center" style="display: flex; justify-content: center; align-items: center;"> <img src="stepfun.svg" width="25" style="margin-right: 10px;"/> <h1 style="margin: 0; border-bottom: none;">Step 3.5 Flash Base Midtrain</h1> </div>

</div>

1. Introduction

Step 3.5 Flash (visit website) is our most capable open-source foundation model, engineered to deliver frontier reasoning and agentic capabilities with exceptional efficiency. We also open-sourced the training codebase (SteptronOss), with support for continue pretrain, SFT, RL (WIP), and evaluation (WIP), and will open-source the SFT data. Built on a sparse Mixture of Experts (MoE) architecture, it selectively activates only 11B of its 196B parameters per token. This "intelligence density" allows it to rival the reasoning depth of top-tier proprietary models, while maintaining the agility required for real-time interaction.

2. Key Capabilities

• Deep Reasoning at Speed: While chatbots are built for reading, agents must reason fast. Powered by 3-way Multi-Token Prediction (MTP-3), Step 3.5 Flash achieves a generation throughput of 100–300 tok/s in typical usage (peaking at 350 tok/s for single-stream coding tasks). This allows for complex, multi-step reasoning chains with immediate responsiveness.

• A Robust Engine for Coding & Agents: Step 3.5 Flash is purpose-built for agentic tasks, integrating a scalable RL framework that drives consistent self-improvement. It achieves 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, proving its ability to handle sophisticated, long-horizon tasks with unwavering stability.

• Efficient Long Context: The model supports a cost-efficient 256K context window by employing a 3:1 Sliding Window Attention (SWA) ratio—integrating three SWA layers for every full-attention layer. This hybrid approach ensures consistent performance across massive datasets or long codebases while significantly reducing the computational overhead typical of standard long-context models.

• Accessible Local Deployment: Optimized for accessibility, Step 3.5 Flash brings elite-level intelligence to local environments. It runs securely on high-end consumer hardware (e.g., Mac Studio M4 Max, NVIDIA DGX Spark), ensuring data privacy without sacrificing performance.

3. Performance

Step 3.5 Flash delivers performance parity with leading closed-source systems while remaining open and efficient.

Performance of Step 3.5 Flash measured across Reasoning, Coding, and Agentic Abilities. Open-source models (left) are sorted by their total parameter count, while top-tier proprietary models are shown on the right. xbench-DeepSearch scores are sourced from official publications for consistency. The shadowed bars represent the enhanced performance of Step 3.5 Flash using Parallel Thinking.

Detailed Benchmarks

1. “*” denotes cases where the original score was unavailable; we report results evaluated under the same test conditions as Step3.5 Flash for fair comparison.
2. “†” indicates DeepSeek scores quoted from the MiMo‑V2‑Flash report.

Recommended Inference Parameters

1. For general chat domain, we suggest: temperature=0.6, top_p=0.95
2. For reasoning / agent scenario, we recommend: temperature=1.0, top_p=0.95.

4. Architecture Details

Step 3.5 Flash is built on a Sparse Mixture-of-Experts (MoE) transformer architecture, optimized for high throughput and low VRAM usage during inference.

4.1 Technical Specifications

4.2 Mixture of Experts (MoE) Routing

Unlike traditional dense models, Step 3.5 Flash uses a fine-grained routing strategy to maximize efficiency:

• Fine-Grained Experts: 288 routed experts per layer + 1 shared expert (always active).
• Sparse Activation: Only the Top-8 experts are selected per token.
• Result: The model retains the "memory" of a 196B parameter model but executes with the speed of an 11B model.

4.3 Multi-Token Prediction (MTP)

To improve inference speed, we utilize a specialized MTP Head consisting of a sliding-window attention mechanism and a dense Feed-Forward Network (FFN). This module predicts 4 tokens simultaneously in a single forward pass, significantly accelerating inference without degrading quality.

5. Training Codebase

The training codebase for Step 3.5 Flash is available at SteptronOss.

📜 Citation

If you find this project useful in your research, please cite our technical report:

@misc{huang2026step35flashopen,
      title={Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters}, 
      author={Ailin Huang and Ang Li and Aobo Kong and Bin Wang and Binxing Jiao and Bo Dong and Bojun Wang and Boyu Chen and Brian Li and Buyun Ma and Chang Su and Changxin Miao and Changyi Wan and Chao Lou and Chen Hu and Chen Xu and Chenfeng Yu and Chengting Feng and Chengyuan Yao and Chunrui Han and Dan Ma and Dapeng Shi and Daxin Jiang and Dehua Ma and Deshan Sun and Di Qi and Enle Liu and Fajie Zhang and Fanqi Wan and Guanzhe Huang and Gulin Yan and Guoliang Cao and Guopeng Li and Han Cheng and Hangyu Guo and Hanshan Zhang and Hao Nie and Haonan Jia and Haoran Lv and Hebin Zhou and Hekun Lv and Heng Wang and Heung-Yeung Shum and Hongbo Huang and Hongbo Peng and Hongyu Zhou and Hongyuan Wang and Houyong Chen and Huangxi Zhu and Huimin Wu and Huiyong Guo and Jia Wang and Jian Zhou and Jianjian Sun and Jiaoren Wu and Jiaran Zhang and Jiashu Lv and Jiashuo Liu and Jiayi Fu and Jiayu Liu and Jie Cheng and Jie Luo and Jie Yang and Jie Zhou and Jieyi Hou and Jing Bai and Jingcheng Hu and Jingjing Xie and Jingwei Wu and Jingyang Zhang and Jishi Zhou and Junfeng Liu and Junzhe Lin and Ka Man Lo and Kai Liang and Kaibo Liu and Kaijun Tan and Kaiwen Yan and Kaixiang Li and Kang An and Kangheng Lin and Lei Yang and Liang Lv and Liang Zhao and Liangyu Chen and Lieyu Shi and Liguo Tan and Lin Lin and Lina Chen and Luck Ma and Mengqiang Ren and Michael Li and Ming Li and Mingliang Li and Mingming Zhang and Mingrui Chen and Mitt Huang and Na Wang and Peng Liu and Qi Han and Qian Zhao and Qinglin He and Qinxin Du and Qiuping Wu and Quan Sun and Rongqiu Yang and Ruihang Miao and Ruixin Han and Ruosi Wan and Ruyan Guo and Shan Wang and Shaoliang Pang and Shaowen Yang and Shengjie Fan and Shijie Shang and Shiliang Yang and Shiwei Li and Shuangshuang Tian and Siqi Liu and Siye Wu and Siyu Chen and Song Yuan and Tiancheng Cao and Tianchi Yue and Tianhao Cheng and Tianning Li and Tingdan Luo and Wang You and Wei Ji and Wei Yuan and Wei Zhang and Weibo Wu and Weihao Xie and Wen Sun and Wenjin Deng and Wenzhen Zheng and Wuxun Xie and Xiangfeng Wang and Xiangwen Kong and Xiangyu Liu and Xiangyu Zhang and Xiaobo Yang and Xiaojia Liu and Xiaolan Yuan and Xiaoran Jiao and Xiaoxiao Ren and Xiaoyun Zhang and Xin Li and Xin Liu and Xin Wu and Xing Chen and Xingping Yang and Xinran Wang and Xu Zhao and Xuan He and Xuanti Feng and Xuedan Cai and Xuqiang Zhou and Yanbo Yu and Yang Li and Yang Xu and Yanlin Lai and Yanming Xu and Yaoyu Wang and Yeqing Shen and Yibo Zhu and Yichen Lv and Yicheng Cao and Yifeng Gong and Yijing Yang and Yikun Yang and Yin Zhao and Yingxiu Zhao and Yinmin Zhang and Yitong Zhang and Yixuan Zhang and Yiyang Chen and Yongchi Zhao and Yongshen Long and Yongyao Wang and Yousong Guan and Yu Zhou and Yuang Peng and Yuanhao Ding and Yuantao Fan and Yuanzhen Yang and Yuchu Luo and Yudi Zhao and Yue Peng and Yueqiang Lin and Yufan Lu and Yuling Zhao and Yunzhou Ju and Yurong Zhang and Yusheng Li and Yuxiang Yang and Yuyang Chen and Yuzhu Cai and Zejia Weng and Zetao Hong and Zexi Li and Zhe Xie and Zheng Ge and Zheng Gong and Zheng Zeng and Zhenyi Lu and Zhewei Huang and Zhichao Chang and Zhiguo Huang and Zhiheng Hu and Zidong Yang and Zili Wang and Ziqi Ren and Zixin Zhang and Zixuan Wang},
      year={2026},
      eprint={2602.10604},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.10604}, 
}

License

This project is open-sourced under the Apache 2.0 License.