Stepfun

Listed at https://platform.stepfun.com

Overall Rank #6 ⭐⭐ Recommended
✅ Direct from China (platform.stepfun.com endpoint, domestic BGP routes) | 🇨🇳 China Direct

💰 Token Pricing

TypePriceNote
Input Step-2: ¥6/M, Step-2-mini: ¥1/M, Step-1: ¥4/M, Step-R: ¥8/M per million tokens
Output Step-2: ¥18/M, Step-2-mini: ¥3/M, Step-1: ¥12/M, Step-R: ¥24/M per million tokens
💡 Free Credits: New accounts receive 1,000,000 free tokens (30-day validity from registration)

🤖 Supported Models (10)

Step-2 (1.2T multimodal flagship, text+image+audio+video, 128K context)Step-2-mini (200B multimodal, faster Step-2 variant)Step-1 (300B text+image multimodal)Step-1V (vision-tuned, 100B)Step-1.5V (multimodal upgrade, 200B)Step-1o (omni multimodal, audio+video)Step-R (reasoning, o1/R1 class)Step-CC (code generation, completions API)Step-Embed (text embedding, 1024-d)Step-Rerank (bge-style reranker)

✨ Pros

  • Step-2 trillion-parameter multimodal flagship: text+image+audio+video on a single endpoint
  • Direct China access with no proxy, typical TTFT 200-400ms
  • OpenAI-compatible /v1/chat/completions endpoint — one-line base_url migration
  • Step-R reasoning model at 1/5 the price of OpenAI o1 / DeepSeek R1
  • Step-CC code completion with IDE-friendly completions API

⚠️ Cons

  • ×Step-2 context window is 128K, behind Kimi K3 (1M) and Qwen3.8-Max (1M)
  • ×International endpoint limited to Hong Kong, higher latency from US/EU
  • ×No first-class Function Calling support — must simulate via prompts
  • ×Model release cadence is moderate, GA timeline for new models is long
  • ×No special discounts beyond volume credits

🎯 Best For

Domestic multimodal production workloads; agents that need single-endpoint image/video understanding; cost-sensitive Step-R reasoning as 1/5-price o1 alternative

💰 Pricing & Plans

ModelInput (¥/M tokens)Output (¥/M tokens)Notes
Step-2¥6 (≈$0.83)¥18 (≈$2.50)Trillion-param multimodal flagship, 128K context
Step-2-mini¥1 (≈$0.14)¥3 (≈$0.42)Fast Step-2 variant, 32K context, 200B params
Step-1¥4 (≈$0.56)¥12 (≈$1.67)300B multimodal, 32K context, prior gen
Step-1V / Step-1.5V¥3 (≈$0.42)¥9 (≈$1.25)Vision-tuned, 100B-200B
Step-R¥8 (≈$1.11)¥24 (≈$3.33)Reasoning (o1 / DeepSeek R1 class)
Step-CC¥1 (≈$0.14)¥2 (≈$0.28)Code completion, IDE-friendly

🔧 API & Developer Experience

  • API Style: OpenAI-compatible /v1/chat/completions endpoint — drop-in replacement by changing base_url.
  • Base URL: https://platform.stepfun.com/v1 — direct from mainland China, no proxy required.
  • SDK Compatibility: Works with the official OpenAI Python/Node SDKs by swapping base_url and api_key. Stepfun also ships a first-party Python SDK with multimodal helpers at github.com/stepfun-ai.
  • Multimodal Inputs: Step-2 and Step-2-mini accept image_url, video_url and audio_url content parts in the same chat-completions request — no separate vision API.
  • Streaming: Server-Sent Events (SSE) streaming for both text and multimodal responses; token usage is reported in the final chunk.
  • Function Calling: Beta: tools parameter accepted, but structured tool_choice is incomplete. Production callers usually wrap Step-2 in a prompt-engineering layer for tool routing.
  • Rate Limits: Default 60 RPM / 100,000 TPM on free tier, raises to 600 RPM / 10M TPM after first top-up. Concurrency 5/20/100 across tier ladder.

🧠 Step-2 万亿参数多模态能力

Step-2 is Stepfun's 1.2T-parameter multimodal flagship, accepting text, image, audio and video inputs through a single /v1/chat/completions surface. Where competing China LLMs (DeepSeek V3.2, Qwen3.5-Max, Moonshot Kimi K3) require separate vision endpoints or per-modality pricing, Step-2 treats every input as a content part, so a video understanding call is billed the same way as a text chat. The 128K context window is mid-pack — behind Kimi K3 (1M) and Qwen3.8-Max (1M), but ahead of DeepSeek V3.2 (64K default, 128K max). For multimodal benchmarks Step-2 lands inside the top-3 of the China open-weight cohort on MMMU, MMBench and MathVista, and the Step-1.5V vision model ranked top-1 on the Chinese-language CV-Bench-QA in early 2026. In production the model is a strong fit for short-clip video understanding (≤2 minutes), OCR-heavy document reasoning, and long-form image+text RAG — the context window and structured output stability are its biggest advantages over DeepSeek V3.2's text-first pipeline.

🌐 Regional Availability & Latency

Stepfun is the second major China LLM provider (after DeepSeek) with first-party direct access from mainland China. The platform.stepfun.com endpoint serves China-based traffic over domestic BGP routes, with typical TTFT of 200-400ms and generation throughput of 40-80 tokens/sec for Step-2. International traffic is served via the Hong Kong PoP, where US/EU callers see 150-300ms TTFT. There is no published Singapore or Tokyo edge as of August 2026 — agents serving East Asia outside China should benchmark before deploying. For production workloads in mainland China, Stepfun is one of three credible first-party endpoints (alongside DeepSeek and SiliconFlow) that bypass the OpenAI / Anthropic proxy requirement entirely. Cross-region routing is the standard production pattern for agents that need both domestic China and overseas users — Stepfun for the China leg, OpenRouter or the upstream provider directly for the international leg. The OpenAI-compatible surface makes this split trivial: only the base_url changes per request.