AI Daily 2026.11.05: Baidu Open-Sources ERNIE 5.5 Flash Base; DeepSeek Harness Hits 1.54M DAU

Model open source, agent infrastructure, and compute efficiency are advancing together. AI is moving from isolated capabilities to system-level competition.

AI Daily 2026.11.05: Models, Agents, and Compute Infrastructure Accelerate

Baidu Open-Sources ERNIE 5.5 Flash Base

On November 5, Baidu officially open-sourced its latest base model, ERNIE 5.5 Flash Base. The model uses a Mixture-of-Experts architecture with 280 billion total parameters and 14 billion activated per token. It was pretrained on 33.5 trillion tokens, supports up to 128K context during pretraining, and allows up to 1 million tokens during fine-tuning and inference.

This gives developers a high-context Chinese base model that can be deployed locally or in private environments, especially useful for long documents, codebases, and enterprise knowledge bases.

DeepSeek Harness Reaches 1.54M DAU and 92 Trillion Daily Tokens

DeepSeek Harness disclosed that, as of November 4, its official API user DAU reached 1.54 million, with daily token consumption of approximately 92 trillion. The number of third-party plugins grew to around 35,000, and about 60% of active users have created or installed at least one plugin.

This shows rapidly growing developer demand for agent platforms, plugin ecosystems, and API usage. For global teams, maintaining stable multi-region API access, managing costs, and ensuring service continuity will become increasingly important.

NVIDIA Releases Vera CPU Benchmarks for Agent Workloads

NVIDIA published benchmark results for Vera CPU on agent workloads. In a 100-megawatt data center scenario, Vera CPU supports 1.6x more agents per watt than the fastest x86 CPU, and delivers 2x faster agent execution than the most energy-efficient x86 CPU.

As enterprises scale agent deployments, data center bottlenecks are shifting from raw compute to a balance of speed, concurrency, and energy efficiency. Early partners including Perplexity, DeepInfra, Daytona, and Modal reported performance gains in agent coding, orchestration, and sandbox concurrency.

Moonshot AI Launches Hosted Agents

Moonshot AI officially launched hosted agents, with analysts noting that strong agent execution is becoming a key focus of current iterations.

Hosted agents shift the focus from “can the model answer” to “can the system continuously execute, integrate with business tools, handle failures, and close the loop.” This aligns with broader trends around Super-agent architectures and enterprise agent platforms.

Wuzhen Summit: Open Source and AI Agents Take Center Stage

During the 2026 World Internet Conference Wuzhen Summit, 20 organizations including Huawei, Alibaba, Tencent, Xiaomi, and Douyin jointly released the Self-Discipline Convention for Agent Development, covering agent identity, cross-domain trust, and security.

The summit’s blue book reported that China’s enterprise AI agent market reached RMB 21.2 billion in 2025, while the embodied AI market was about RMB 915 billion.

Tencent also showcased Hunyuan Hy4 preview and announced full open source, targeting software engineering, research, and intelligent office scenarios.

What Developers Can Do Next

  • Deploy DeepSeek Harness with one click via Celedog.io and quickly connect to Super-agent and plugin ecosystems.
  • Use Celedog.io’s massive Agent Skill library for coding, data analysis, DevOps automation, content creation, and more.
  • Build multi-model routing and cost-optimization strategies using open models such as ERNIE 5.5 Flash Base and Hunyuan Hy4 preview.
  • For global multi-site operations, prioritize API stability, regional latency, concurrency quotas, and failover mechanisms.

Last updated November 5, 2026

Where to go next