Breaking Dimensions! The SenseCore Behind The Viral App

50% cost reduction, 3-5x lower inference latency, 100% QPS (queries per second) boost!

Zaomeng Ciyuan (造梦次元), a next-generation multimodal AI-powered content platform created by IdeaFlow, is transforming digital experiences. Powered by SenseCore's end-to-end AIGC solutions spanning from computing infrastructure to advanced models, the platform achieves optimal coordination between computing power and models, as well as between models and applications. This seamless integration delivers both exceptional user experiences and maximum cost efficiency. IdeaFlow is a breakout success and has become a popular platform with users averaging over 100 minutes of daily engagement.

Zhang Hong, Partner at IdeaFlow and Head of Studio, stated, "SenseCore demonstrates three core advantages. The first is its deep integration between computing infrastructure and AI models which enables maximally efficient resource utilization. Second, the seamless coordination across text, voice, image and other modalities creates fluid, natural interactive experiences. Third, its robust ecosystem supports open-source models which provide diverse model options. This multi-dimensional synergy has helped us achieve holistic optimization across the entire chain of 'computing power, models, and applications,' driving efficient transformation of technical capabilities into business value."

60 Million Daily API Calls & 100 Billion Token Consumption Pose Three Major "Survival Challenges"

Based on deep user insights, IdeaFlow's technological innovations directly address two core challenges in AI interaction applications: "lack of immersion" and "high barriers to entry".

For example, by providing multimodal interactions combining text, voice, and text-image integration, the user-AI interaction experience approaches realistic human communication. IdeaFlow also offers abundant creation tools where users can simply apply relevant templates to quickly build complete interactive content, effectively lowering the creative threshold.

IdeaFlow handles over 60 million daily model calls on average, and billions of tokens consumed daily. At this massive scale, any lag, downtime, or errors can lead to significant user churn, subpar experiences due to inadequate model capabilities damage product reputation, and minor resource idling accumulating into substantial costs. These are the three key "survival challenges" that IdeaFlow faces:

2x Peak-to-Trough Computing Elasticity Battle: IdeaFlow's platform traffic shows distinct patterns of peaks and troughs, with peak periods (weekends/ holidays) reaching twice the trough volume, demanding highly elastic infrastructure computing resources.

The Critical 2-Second Inference Latency Threshold: Virtual character interactions are very latency-sensitive. To ensure basic smooth interaction, model inference latency must be consistently controlled within 2 seconds.

Stable Model Iteration Challenge: IdeaFlow utilizes many open-source models requiring frequent upgrades or replacements following community rhythms. With numerous demands and frequent updates, maintaining stable operations is crucial.

Lag-Free Peak Performance, Waste-Free Low-Usage Periods

SenseCore, recognized as "the AI infrastructure that best understands large models" provides IdeaFlow with an end-to-end integrated AIGC solution spanning from computing power to models. This one-stop solution comprehensively supports IdeaFlow's product development and operational needs, delivering “Lag-free peak performance, waste-free low-usage periods”

Second-Level Elastic Scaling: 50% Cost Reduction with Seamless Traffic Adaptation

To address peak-valley traffic fluctuations, SenseCore's combined strategy of real-time monitoring, unified scheduling, and intelligent scaling achieves second-level elastic scaling, reducing IdeaFlow's operational costs by 50%. Its intelligent unified scheduling system:

• Automatically allocates resources based on real-time metrics and scaling rules

• Precisely matches computing power to traffic demands

• Enhances flexibility through dual scaling strategies (scheduled + on-demand)

Full-Stack Optimization: 5X Lower Latency & 100% QPS Boost

With "interactions per user" as its core metric, IdeaFlow leveraged SenseCore's multi-dimensional optimizations across hardware, software frameworks, and algorithms to:

• Reduce inference latency by 3-5X

• Implement prioritized traffic routing during peaks

• Achieve 100% QPS improvement through quantized acceleration

SenseTime's Multi-Model Suite Doubles "Interactions Per User"

Another key factor driving "interactions per user" is model capability. SenseTime's integrated model suite—including the SenseChat large language model, SenseChat-Character humanlike dialogue model, and SenseMirage text-to-image model—delivers profoundly intuitive interactions for IdeaFlow users, creating experiences that "truly understand human emotions."

SenseChat enables immersive, human-like conversations through: Precise contextual semantic analysis; Advanced intent reasoning; Emotionally aware responses.

SenseAvatar elevates role-playing with industry-leading capabilities: Character persona consistency; Plot-driven dialogues; Hyper-realistic conversational experiences.

SenseMirage empowers creators with: Standard/Conditional image generation.

Through infrastructure optimizations and joint model development, IdeaFlow achieved 2X growth in interactions per user (from 20 to 40–50 sessions), and breakthrough improvements in user retention.

Comprehensive Stability Assurance: Seamless Model Upgrades, Zero Service Disruption

To address the stability challenges posed by IdeaFlow's high-frequency model switching and frequent version upgrades of open-source models in business operations, SenseCore has built a comprehensive stability assurance system for the inference phase, covering: Pilot release of models; Rolling upgrades; Intelligent operations. This system not only provides effective support for rapid model iteration, but also ensures stable operation of online services.

Daily 100-Min Engagement: Building New Content Consumption Ecosystem

In just two years, IdeaFlow has evolved into a platform featuring: Hundreds of IP characters with 10K+ fan followings and average daily engagement exceeding 100 minutes. It has become an essential companion for young users in: Entertainment; Emotional expression; Learning companionship. Moving forward, both parties will deepen their collaboration on multimodal model capabilities, delivering even more novel and engaging AI interaction experiences.

Yang Fan, Co-founder of SenseTime and President of SenseCore, stated, "The collaboration between SenseCore and IdeaFlow perfectly exemplifies the strategic value of SenseTime's 'AI Infrastructure - Large Models - Applications' integrated strategy. By co-developing model capabilities tailored to specific application scenarios, we train vertical models that precisely meet business needs, while providing optimal infrastructure support for model deployment—including flexible elastic scaling, ultra-low inference latency, and robust stability guarantees. This ultimately achieves both cost efficiency improvements and enhanced user satisfaction. The joint optimization of large model algorithms and infrastructure represents not only the core driver for advancing generative AI, but also the optimal pathway to maximize commercial value."