---
id: "2026-10-15-stepfun-step-5-preview-open-weights"
url: "https://postcutoff.com/e/2026-10-15-stepfun-step-5-preview-open-weights/"
as_of: "2026-10-15T10:09:00+02:00"
date: "2026-10-15"
date_precision: day
category: open-source
importance: 3
confidence: high
status: [Confirmed]
sources: 6
editor: human editor
human_review: null
version: "2026-10-15"
---

As of: 2026-10-15 10:09 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-15-stepfun-step-5-preview-open-weights/

# StepFun releases Step 5 Preview open weights

Full title: StepFun publishes Step 5 Preview open weights on schedule: a 600B MoE agent model under a custom license that treats output-trained models as derivatives

On Oct 15, 2026 (Beijing time) StepFun made the weights of Step 5 Preview public on Hugging Face and ModelScope, as promised at its Sept 20 API launch: a 600B-total / 27B-active sparse MoE with up to 1M tokens of context, in FP8, BF16 and NVFP4 checkpoints. The StepFun Model License 1.0 is Apache-style but requires a commercial licence for device makers and for large model-as-a-service providers, "Powered by StepFun AI" attribution above 100M users or $20M monthly revenue, and counts models trained materially on Step 5's outputs as derivative works.

## Key facts
- Repos (HF API): stepfun-ai/Step-5-Preview (FP8, 24 weight shards plus MTP and vision files), stepfun-ai/Step-5-Preview-BF16 and stepfun-ai/Step-5-Preview-NVFP4; all public and ungated, last modified about 01:30 UTC on Oct 15, 2026; license tag 'other'
- Model card: sparse MoE, 600B total / 27B active parameters, context up to 1M tokens, maximum output up to 1M tokens; the spec table lists text and image input, text output (the introduction also mentions video)
- Positioning (model card): 'StepFun's flagship model for agentic work', with 'particular strength in finance'; benchmark results are shown only as an image; GDPval-AA v2.1 scores are from Artificial Analysis as of Sept 20, 2026
- Serving: vLLM (dedicated docker image vllm/vllm-openai:step-5-preview), SGLang and a standalone Transformers implementation; FP8 fits one node of 4 B300 or 8 H200 GPUs, BF16 needs 8 B300 or 16 H200 GPUs; recommended sampling temperature 1.0, top_p 0.95
- License, StepFun Model License 1.0: perpetual, royalty-free use, modification and redistribution with a patent grant; 'A model trained to a material extent on outputs of the Model is deemed to reproduce or contain a substantial part of the Model' (so such models are derivative works under the licence)
- License conditions: a separate commercial licence is needed (i) by hardware makers (phones, vehicles, smart glasses, wearables, robots, smart-home devices) shipping its capabilities and (ii) by model-as-a-service providers with over $20M revenue in 12 months; products above 100M monthly active users or $20M monthly revenue must display 'Powered by StepFun AI'. Internal use is exempt
- On time: StepFun had promised Oct 15 at launch; on Oct 13–14 the placeholder repo was still private (HTTP 401). API access (step-5-preview, $1.00/$2.70 per 1M tokens) and OpenRouter listing continue

## What happened

StepFun kept the date it set at launch. Shortly after midnight UTC on Oct 15 (morning in Beijing) its three Hugging Face repos
became public with full weight files: the main FP8 checkpoint, a BF16 version and an NVFP4 version for Blackwell GPUs, plus
ModelScope mirrors. The model card repeats the launch specification (600B total / 27B active MoE, 1M context) and gives
deployment recipes for vLLM, SGLang and Transformers.

The weights come with a new licence, the StepFun Model License 1.0. It reads like Apache 2.0 for most users but adds three
commercial gates: device makers need a separate licence before shipping Step 5 capabilities in phones, cars, glasses, robots
or smart-home products; model-as-a-service providers with over $20M of annual revenue need one too; and very large products
must show "Powered by StepFun AI". It also defines any model "trained to a material extent on outputs of the Model" as a
derivative work, which brings distillation from Step 5 under the licence's attribution and commercial terms.

## Why it matters

Step 5 Preview is one of the largest openly downloadable agent models from a Chinese lab, a day after Ant's 560B Ling 3.1
Flash (MIT). The licence's distillation clause is unusual for open weights and mirrors the industry argument over distillation
that Anthropic's Oct 12 report revived. The device clause matters because StepFun's own agent phone (StepX Neo) launched the
same week.

Benchmarks remain vendor-reported (shown only as an image). Press coverage of the release was not yet available at writing.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 168 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 107 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 198 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 137 days after its cutoff

## Sources

1. [Hugging Face: stepfun-ai/Step-5-Preview (model card)](https://huggingface.co/stepfun-ai/Step-5-Preview) (huggingface.co, code)
2. [Hugging Face: stepfun-ai/Step-5-Preview-BF16](https://huggingface.co/stepfun-ai/Step-5-Preview-BF16) (huggingface.co, code)
3. [Hugging Face: stepfun-ai/Step-5-Preview-NVFP4](https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4) (huggingface.co, code)
4. [StepFun Model License 1.0 (LICENSE.txt)](https://huggingface.co/stepfun-ai/Step-5-Preview/blob/main/LICENSE.txt) (huggingface.co, docs)
5. [StepFun docs: Step 5 Preview](https://platform.stepfun.ai/docs/en/guides/models/step-5-preview) (platform.stepfun.ai, docs)
6. [Pandaily: StepFun Step 5 Preview, 600B MoE, open weights Oct 15 (Sept 20)](https://pandaily.com/stepfun-step-5-preview-600b-moe-1m-context-open-weights-oct-15) (pandaily.com, press)

## Changes

- 2026-10-15 (filed): Created (HF API and model card read directly; resolves upcoming item `[StepFun Step 5 Preview open weights promised for Oct 15](https://postcutoff.com/upcoming/receipts/#2026-10-15-step-5-open-weights)`)

## Related

- 2026-10-14: [inclusionAI releases Ling 3.1 Flash open weights](https://postcutoff.com/e/2026-10-14-ling-3-1-flash-open-weights/index.md)
- 2026-10-13: [StepFun's device unit launches the STEPX Neo 'LLM-native agent phone' at 3,999 yuan, bundled with 10 billion tokens a month for a year](https://postcutoff.com/e/2026-10-13-stepfun-stepx-neo-agent-phone-launch/index.md)
- 2026-10-12: [Anthropic says new defenses cut the Claude reasoning harvested by distillers by about 90 to 99 percent](https://postcutoff.com/e/2026-10-12-anthropic-distillation-defenses/index.md)
- 2026-09-20: [StepFun launches Step 5 Preview, a 600B-parameter MoE agent model with 1M context](https://postcutoff.com/e/2026-09-20-stepfun-step-5-preview/index.md)