Stepfun Low-Inference-Cost Document and Diagram

Strengths

• Trillion-parameter MoE architecture excels in complex reasoning tasks
• Step 3.7 Flash achieves an inference speed of 400-409 Tokens/s, driving inference costs down to edge-device levels
• Open-source models maintain strong competitiveness in lightweight domains, ranking among the top in call volume across multiple open platforms

Weaknesses

• Flagship model has a large parameter size, leading to higher time-to-first-token (TTFT) and impacting interactive real-time performance
• Prone to hallucinations, occasionally and confidently fabricating non-existent function names, which affects reliability in coding scenarios
• Operating for less than three years, making it weaker than leading vendors in brand trust and large-scale enterprise deployment case accumulation

Opportunities

• Global computing constraints drive up cloud costs, making low-inference-cost models more price-competitive
• Enterprise-grade document intelligence and code-assistance demand continues to grow as AI applications deepen
• Plans for an IPO in Hong Kong provide capital backing, expected to enhance enterprise customer acquisition capabilities

Threats

• Domestic vendors such as DeepSeek, MiniMax, and Zhipu are closing in on inference costs and scenario capabilities
• Continuous iterations of open-source models like Meta Llama and Mistral put downward pressure on the pricing structure
• Computing resource constraints and rising cloud service costs compress profit margins