According to Sina Finance, the number of developers on the Zhipu MaaS (Model as a Service) platform has exceeded 1000 ten thousand. On actual model usage, based on official statistics from US-based Databricks (as of the 11th month, 2th day), measured by token usage in the open-weight model category, the Zhipu GLM series accounts for over 60%. This data reflects the invocation scale of its open-weight models on third-party platforms; specific statistical methodology is subject to the original scope.
Doubao officially announced on 2026-11-18 that Doubao Work, an office product launched three months earlier, surpassed 8000,000 daily active users (DAU), with enterprise customer annual recurring revenue (ARR) exceeding RMB 2000 million. This is the first batch of scaled operational data publicly disclosed for the product, reflecting its user growth in work scenarios and progress in enterprise monetization.
Volcano Engine officially launched the API service of its Doubao real-time audio-video conversation model (SeedRealtime 1.0) on 11 19. The model unifies audio, visuals, temporal sequencing, and expression into a single end-to-end model, supporting "watch, listen, and speak simultaneously" full-modal real-time interaction. Compared with traditional cascade solutions chaining ASR, VLM, and TTS, it reduces accumulated latency and information loss, and features continuous environmental awareness, proactive expression, and noise resilience.
At the Baidu World 2026 conference on 11 12, Baidu released the ERNIE Lightning deep-thinking model and the ERNIE Yijing micro-expression large model, both launched on the Qianfan platform with API open access; it also disclosed plans to launch Kunlunxin M300 in 2027 and the Tianchi 1024-card supernode in 2027, along with the G3.5 KV Cache acceleration cabinet. The overall roadmap centers on reducing inference costs and scaling inference.
On 2026 11 1, the Yangtze River Cup 2026 Global AI Protein Design Challenge announced seven targets and competition rules, planning to provide free wet-lab validation for approximately 7000 AI de novo-designed candidate proteins, with rankings based on officially measured KD. Within one month of registration opening, more than 1800 teams and over 4000 participants had joined, covering more than 20 countries and regions; registration and sequence submission close on 2026 11 30.
In 'Ten Trends in Emerging-Domain and New-Quality Innovation-Driven Development,' the China Center for Information Industry Development (CCID) noted that China's AI inference data volume has for the first time exceeded its training data volume. As of 2026 6, the national daily average token usage surpassed 175 trillion, up more than 1000-fold from the start of 2024. The article also cites data on smart factory construction, industry application deployment, and generative AI user scale, showing that AI is shifting from model training to large-scale industrial use.
The People's Forum published an article by Wang Feiyue, researcher at the Institute of Automation, Chinese Academy of Sciences, reviewing the industrial adoption and competitiveness of Chinese open-weight models on platforms such as Hugging Face and OpenRouter, covering overseas commercial cases of models including KimiK3, GLM-5.3, DeepSeek, and Qwen3-VL; it also points out risks such as malicious fine-tuning and deepfakes, and explores paths from open technology to open governance and organizational intelligence.
Tech Corner has released a speaker preview for Hongqiao International Tech Week: Ning Ning, Vice President of International Business at 01.AI (01.AI), will deliver a keynote speech titled "Building an AI Native Future: Cross-Industry Practices and Insights" on the AI Stage on 22, 11, 2026. Drawing on her global implementation experience across education, industry, government, and entrepreneurship, she will explore how enterprises can transform from AI-enabled to AI Native organizations.
On 11/12, iFLYTEK announced that the iFLYTEK Spark Speech Foundation Large Model SparkAudio1.0 has launched the API service on the iFLYTEK Open Platform, and developers can integrate and call it directly. The model supports dual-modality input (text and speech) with text output, covering tasks such as speech transcription, multilingual translation, multi-dialect recognition, and speaker separation. Officially, it supports 99 languages and 202 dialects and is trained on fully domestically produced compute.
In a research report published on 2026, 11 16, Shanxi Securities maintained a Buy rating on Kunlun Tech (300418). The report disclosed that the company's revenue for the first three quarters of 2026 was RMB 89.37 hundred million, up 53.96% YoY, with net profit attributable to parent of RMB 260.94 ten thousand, turning from loss to profit. Tiangong AI business revenue was RMB 43.21 hundred million, up 168.29% YoY, and the AI short-drama platform generated nearly USD 1 hundred million in monthly run-rate in 26 9. Brokerages forecast net profit attributable to parent of RMB -3.59, 0.94, and 5.39 hundred million for 2026–2028.
ByteDance's Seed team released the image generation model Seedream 5.5 on 2026 11, 9, as the successor to Seedream 5.0 Pro. The new model focuses on upgrades in multi-turn editing consistency, native RGBA transparent image output, and photorealistic quality, while improving accuracy in rendering complex Chinese and English glyphs and dense small text. Official evaluations claim the usable rate across six scenario categories improves over the previous generation, with film/TV and office scenarios more than doubling. The model is being rolled out progressively on Jimeng, Doubao, Doubao Work, and Volcano Ark; API pricing has not yet been announced.