TL;DR: Stepfun (阶跃星辰) has the largest model (1T+ params) and best multimodal capabilities. DeepSeek offers the best reasoning-to-cost ratio with open-source models. Kimi (月之暗面) excels at ultra-long context (200万字) and agent capabilities.
Feature Comparison
| Feature | Stepfun (阶跃星辰) | DeepSeek | Kimi (月之暗面) |
|---|---|---|---|
| Funding | $718M (Jan 2026) | Well-funded | Well-funded |
| Model Size | 1T+ parameters | Efficient (MoE) | Large |
| Modalities | Text, image, video, audio | Text, code, reasoning | Text, image, video, agent |
| Context Window | Long | 128K | 200万字 (longest) |
| Reasoning | Strong | Best (R1) | Strong |
| Agent Capabilities | Growing | Limited | 100+ parallel agents |
| Open Source | Yes (11 models) | Yes (full weights) | No |
| API Cost | Competitive | Cheapest | Competitive |
| Device Partnerships | Honor, Oppo, ZTE | None | None |
Who Should Choose What
- Stepfun: Multimodal projects needing text + image + video + audio, on-device AI
- DeepSeek: Developers wanting the cheapest high-quality reasoning, open-source community
- Kimi: Long-document analysis, agentic workflows, 200万字 context needs
The "Six Tigers" of Chinese AI
Stepfun, DeepSeek, and Kimi (Moonshot AI) are three of China's "六小虎" (Six Little Tigers) — the most promising AI startups challenging tech giants. They represent different approaches: Stepfun bets on scale, DeepSeek on efficiency, and Kimi on context length and agents.
Frequently Asked Questions
Which is most like ChatGPT?
Kimi's chat interface is closest to the ChatGPT experience, with the added advantage of processing documents up to 200万字.
Which is cheapest for API usage?
DeepSeek, by a significant margin. Their MoE architecture makes inference much cheaper than dense models.
Can I self-host any of these?
DeepSeek and Stepfun offer open-weight models you can self-host. Kimi is currently API-only.
Final Verdict
Best Multimodal: Stepfun
Best Value/Reasoning: DeepSeek
Best Long Context: Kimi
Best Open Source: DeepSeek
