StepFun has released Step 5 Preview as open weights, making it available on Hugging Face and ModelScope. This model is a sparse Mixture-of-Experts architecture designed for long-horizon agentic tasks.
Key points
- StepFun released Step 5 Preview as open weights.
- The model has 600B total parameters and 27B active per token.
- It supports a 1M token context window for agentic tasks.
- Performance is comparable to GLM-5.3 and Kimi K3 on coding.
The Step 5 Preview model boasts 600 billion total parameters, with 27 billion active parameters per token, and supports a context window of up to 1 million tokens. It is intended for developers to self-host for applications such as coding agents, document analysis, and finance workflows.
The architecture features a depth of 92 transformer blocks and utilizes a mix of gated sliding-window attention and full attention layers. Sparse GQA allows full-attention layers to select 4,096 history tokens per query, and speculative decoding with an MTP-3 head provides an approximate 30% end-to-end speedup.
By the numbers
- Total parameters
- 600B
- Active parameters per token
- 27B
- Token context window
- 1M
Performance evaluations place Step 5 Preview near open-weight peers like GLM-5.3 and Kimi K3 on coding tasks, though behind models such as GPT-6 Astra and Claude Opus 5. It achieved 88.3% on the AA-LCR long-context reasoning benchmark but scored 33.3% on Terminal-Bench 4.0.
Sources and references
Original report
StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Modelhuggingface.co/stepfun-ai/Step-5-Preview (opens in a new tab)
- Modelhuggingface.co/stepfun-ai/Step-5-Preview-BF16 (opens in a new tab)
- Linkmodelscope.ai/models/stepfun-ai/Step-5-Preview (opens in a new tab)
- Linkcellcog.ai/blog/step-5-preview (opens in a new tab)
- Linkplatform.stepfun.ai/blog/inside-step-5-preview (opens in a new tab)
- Linkopenrouter.ai/stepfun/step-5-preview (opens in a new tab)
Ref 00CO47VF
Cite this
APA
Interestana AI Editorial. (2026, October 17). StepFun Releases 600B Parameter Open Weights Model. Interestana. https://interestana.com/articles/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context-00co47vf
MLA
“StepFun Releases 600B Parameter Open Weights Model.” Interestana, 17 Oct. 2026, interestana.com/articles/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context-00co47vf.
Chicago
Interestana AI Editorial. “StepFun Releases 600B Parameter Open Weights Model.” Interestana, October 17, 2026. https://interestana.com/articles/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context-00co47vf.
BibTeX
@misc{interestana_00co47vf,
author = {{Interestana AI Editorial}},
title = {{StepFun Releases 600B Parameter Open Weights Model}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context-00co47vf},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/16/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context/}
}Credit the source
Original reporting: MarkTechPost, “StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context”, https://www.marktechpost.com/2026/10/16/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe