Baidu ERNIE Qianfan Deployment: Public Cloud Token Pricing vs Private Compute Leasing – Buyer Worksheet Included

Baidu ERNIE and Qianfan Platform Overview Baidu's ERNIE (Enhanced Representation through kNowledge IntEgration) family of models powers the Wenxin (Ask Wenxin) ecosystem, offering industrial-grade, multimodal large language models optimized for enterprise workloads. As of May 13, 2026, Qianfan, the Baidu Cloud platform for ERNIE deployment, supports both public cloud APIs and private on-premises or VPC deployments. This makes it a strong choice for retrieval-augmented generation (RAG) pipelines and AI agents, particularly in regulated sectors. Key models include ERNIE-5.0, the full-modality flagship featuring a Mixture-of-Experts (MoE) architecture designed for advanced reasoning and vision tasks. For scenarios requiring lower latency inference, lighter variants like ERNIE-Lite-8K are available. Qianfan also integrates open-source tools such as ERNIEKit for fine-tuning and FastDeploy for

optimized model serving, providing a seamless bridge from public APIs to private environments. This flexibility is ideal for business-to-business (B2B) operations that require scalable AI solutions without vendor lock-in, especially when dealing with China-centric data flows. Public Cloud Metering: Token Pricing and QPS Limits Qianfan's public cloud service operates on a pay-as-you-go model, with billing based on tokens. Costs are deducted hourly, calculated from the sum of input and output tokens. According to Baidu Cloud Qianfan pricing documentation (cloud.baidu.com/product/qianfan/pricing, as of May 13, 2026), rates vary depending on the specific model SKU: ERNIE-Bot-4.5 : ¥0.024 per 1,000 tokens (input and output combined). ERNIE-5.0 : ¥0.032 per 1,000 tokens for multimodal queries. ERNIE-Lite : ¥0.008 per 1,000 tokens, suitable for cost-sensitive applications. For instance, a RAG

agent handling 10,000 queries daily, with an average of 2,000 input tokens and 500 output tokens per query, would incur approximately ¥6.40 per day when using ERNIE-5.0 (25.5 million tokens \ ¥0.032/1,000 tokens). Additional charges may apply for: RAG Connectors : ¥0.001 per 1,000 retrieved chunks. Image/Video Tokens : These are subject to multipliers (e.g., one image is equivalent to 1,000 tokens). Regional Surcharges : An additional 10-20% may be applied for regions outside mainland China. Query Per Second (QPS) limits for public cloud services typically start at 1 for custom services. These limits can be scaled based on purchased compute units; for example, ERNIE-Bot-turbo supports 10 QPS per unit. Exceeding these limits will result in throttling. Upgrades can be managed through the console. Private Deployment: Compute Leasing Costs and Setup For organizations with high-volume usage r

equirements or specific data sovereignty needs, Qianfan offers private deployment options through compute leasing on Baidu Cloud's infrastructure. This model operates on a rental basis, with costs calculated as: daily rate per compute unit \ number of units \ duration of deployment. Based on official documentation (cloud.baidu.com/product/bfcloud/qianfan private, as of May 13, 2026): A base compute unit (comparable to V100 performance) costs ¥250 per day. For example, deploying 4 units for ERNIE-5.0 serving over 5 days would total ¥5,000. The setup process involves using ERNIEKit for model conversion and fine-tuning, and FastDeploy for optimizing models using TensorRT and ONNX, aiming for latency below 100 milliseconds. The trade-offs for private deployment include: Advantages : Unlimited QPS, the ability to use custom hardware (such as H100 clusters), and no per-token metering. Disadvan

tages : An upfront setup period of 1-2 weeks, and fixed costs regardless of actual utilization. Private deployments are well-suited for steady workloads, such as internal AI agents, and offer Virtual Private Cloud (VPC) isolation for enhanced data residency. Public vs Private: Cost, Latency, and Scalability Comparison The decision between using public token-based metering and private compute leasing depends heavily on workload characteristics: Cost : The public cloud model is more cost-effective for bursty or low-volume usage. For example, at ¥0.032 per 1,000 tokens, the breakeven point for private deployment (at ¥250/day/unit) is around 7.8 million tokens per day. Private deployment becomes more economical for sustained workloads exceeding approximately 100 QPS, as it avoids per-token fees. Latency : Public cloud APIs typically offer an average latency of 200-500 milliseconds, which can

be affected by QPS limits. Private deployments, especially when optimized with FastDeploy on dedicated GPUs, can achieve latencies between 50-150 milliseconds. Scalability : Public cloud services can automatically scale to handle over 1,000 QPS through reservations. Private deployments require manu