StepFun: The "Ex-Microsoft" AI Lab That Chose Independence Over Partnership
In the AI gold rush of 2026, most Chinese AI startups proudly flaunt their Silicon Valley backing. But StepFun—Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd—is writing a different story. Founded in April 2023 by three former Microsoft engineers, this "AI Tiger" company has zero corporate ties to its founders' alma mater. It's a deliberate bet on independence in an industry that loves partnerships.
Let's unpack what that really means for builders.
The Microsoft Alumni Myth
StepFun's founding team reads like a Microsoft AI reunion:
- Jiang Daxin (CEO): 16-year veteran, former Global VP & Chief Scientist at Microsoft Research Asia. Led work on Bing, Cortana, and Azure cognitive services. The Wire China
- Jiao Binxing: Co-founder, ran Microsoft's Search division. The Wire China
- Zhu Yibo: Co-founder, headed Systems engineering. The Wire China
These aren't just engineers—they're the people who built Microsoft's AI infrastructure. Yet when they launched StepFun, they didn't take Microsoft money. No strategic investment. No cloud partnership. No joint go-to-market. Just a clean break.
The reason matters: Microsoft's AI portfolio today includes exclusive partnerships with OpenAI, Mistral, and Meta's Llama. StepFun would have been another supplier—but they chose to compete instead.
No Partnership Means No Safety Net
When you see "ex-Microsoft founders," it's natural to assume Microsoft's backing. The reality is the opposite:
Investors are entirely Chinese: Tencent, Qiming Venture Partners (early ByteDance backer), 5Y Capital, Shanghai State-owned Capital Investment, China Life Private Equity, and Fortera Capital. Their latest Series B+ (January 2026) raised over $718 million. [Research data]
Infrastructure is independent: StepFun runs on its own API platform (api.stepfun.com), not Azure. They're available on OpenRouter and have open-source models on Hugging Face—but no Azure Marketplace listing.
Compliance is self-managed: Unlike Microsoft's partners (OpenAI, etc.) who inherit Microsoft's enterprise certifications (ISO 27001, SOC 2, FedRAMP), StepFun operates on its own compliance framework.
For enterprise buyers, that's a double-edged sword: more flexibility, but also more due diligence.
The 70% Deep Search Proof Point
StepFun's independence isn't just philosophical—it's technical. Their Deep Research feature uses multi-agent architecture to achieve a 70% pass rate on xBench-DeepSearch, outperforming all other models in complex Chinese internet information retrieval. StepFun Deep Research
Why does that matter for global builders?
- It proves they can build production-grade agent systems without Microsoft's Azure AI stack.
- Their models are specialized for complex search workflows—exactly what modern AI agents need.
- It demonstrates technical execution despite being a 3-year-old company.
Their flagship Step-3.5-Flash (196B params, 11B active MoE, 256K context) also scores 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, competitive with GPT-5.2 on coding tasks. [Perplexity research via GitHub/NVIDIA NIM]
Builder Verdict: Independent or Isolated?
Should Western enterprises bet on a China-based, ex-Microsoft, independent AI provider?
The Case For:
- Cost advantage: Step-3.5-Flash is ~5x cheaper than GPT-5.4, with comparable tool-use performance. OpenRouter pricing
- No Microsoft lock-in: You can self-host their open weights or use any cloud provider.
- Proven scale: 11 models on Hugging Face, 70% benchmark pass rates, $718M+ funding.
The Case Against:
- Compliance burden: You must validate their security stack yourself (no Microsoft certifications).
- Geopolitical risk: US-China AI export controls could restrict their model weights or API access.
- Support uncertainty: Without Microsoft's enterprise SLAs, you're dependent on StepFun's direct support.
The Strategic Bet
StepFun's path represents a growing trend: elite AI talent leaving big tech to build independent stacks. The Microsoft alumni aren't trying to be Microsoft's next AI partner—they're trying to be the alternative.
For builders, that means more options. It also means more complexity. But in a world where GPT-5.4's tool-use errors are costing teams productivity, a validated, cost-effective, independent model like Step-3.5-Flash deserves a POC.
The question isn't whether they could have partnered with Microsoft. The question is whether they needed to. Their early results suggest maybe not.
Tags: AI, StepFun, Microsoft, LLM, Builder, Tool-Use, Independence, Chinese AI
