ZanSpark Daily AI Briefing

March 31, 2026
Covering Key Developments (March 30)

1. AI Industry Hot News

【Alibaba Launches Full-Modality Model Qwen3.5-Omni — 215 SOTA Results Surpass Gemini-3.1 Pro】

【Content】 On March 30, Alibaba officially released Qwen3.5-Omni, the next-generation all-modality large model from its Qwen (Tongyi) series. The model supports native multimodal input and output across text, image, audio, and video, employs a hybrid attention MoE architecture, and was pretrained on over 100 million hours of audio-video data. It achieved SOTA on 215 audio-video understanding, recognition, and interaction tasks, comprehensively surpassing Google Gemini-3.1 Pro. Notable highlights include: support for 113 language/dialect speech recognition (including Maori and Hainanese dialect), fine-grained video chapter segmentation with timestamp annotation, camera-based voice-to-code capability for generating app/web/game prototype code directly (Vibe Coding), and API cost of only ~0.8 RMB per million tokens—approximately one-tenth that of Gemini-3.1 Pro.

Key Takeaway: The dual impact of full modality plus ultra-low pricing puts competitive pressure across the board. Short-form video platforms, content moderation systems, and live streaming platforms are gaining a major new capability tier.

🔗 Quantum Bit

【Microsoft Copilot Introduces Multi-Model Collaborative Deep Research: GPT and Claude Working Side by Side】

【Content】 On March 30 (ET), Microsoft delivered a major deep research update for Microsoft 365 Copilot featuring two core capabilities: ① “Critique”—GPT generates an initial research draft, then Claude reviews it against academic standards for accuracy, completeness, and citation quality before producing a comprehensive final report; ② “Council”—multiple models (GPT, Claude, etc.) conduct independent research in parallel, with a “referee model” synthesizing consensus and divergence while summarizing each model’s unique contribution. Microsoft reports that on the DRACO benchmark, multi-model collaboration significantly outperforms any single model while effectively suppressing AI hallucinations. Microsoft stock rose nearly 2% on the day.

Key Takeaway: First-time implementation of competing models working in collaboration. Multi-AI collaboration is emerging as a new enterprise productivity paradigm, with Copilot penetrating professional research scenarios across the board.

🔗 CLS

【Chinese Models Outpace U.S. in Token Usage for Four Consecutive Weeks — Xiaomi MiMo-V2-Pro Retains Global #1】

【Content】 According to latest data from OpenRouter, the world’s largest AI model API aggregation platform, global total AI token usage reached 22.7 trillion tokens during the week of March 23–29, up 11.2% week-over-week. Chinese models once again surpassed U.S. models, marking four consecutive weeks of leadership. Specific rankings: Xiaomi MiMo-V2-Pro retained the top spot with weekly usage of 3.96 trillion tokens; Step-Flash (Step 3.5 Flash) ranked second at 1.49 trillion tokens; MiniMax M2.5 placed ninth after a significant drop from the prior week. The global top four positions are all held by Chinese AI models.

Key Takeaway: Chinese open-source models continue to strengthen their global competitiveness. Xiaomi achieved a 71.2% compute cost reduction through lightweight inference modeling, representing an important pathway toward AI democratization.

🔗 CNMO Tech

【DeepSeek Experiences Two Outages Totaling Nearly 12 Hours — Stability Concerns Emerge for Top-Tier Domestic Models】

【Content】 Beginning the evening of March 29, DeepSeek’s official app and web platform experienced widespread “performance anomalies” preventing normal user responses. The first outage occurred at 21:35 on March 29 and was resolved by 23:23; however, a second anomaly struck at 00:20 on March 30, persisting until after 9:00 AM before gradually recovering—for a cumulative downtime of nearly 12 hours. Notably, integrated applications running DeepSeek models (such as Tencent Yuanbao) were unaffected, indicating that underlying model capabilities remained intact and the issues stemmed from DeepSeek’s own platform architecture and traffic handling capacity.

Key Takeaway: Consecutive outages during a high-profile period expose infrastructure weaknesses among leading domestic models. For enterprise users, service stability will become a critical weighting factor in model selection.

🔗 Sina

2. AI Innovation & Applications

【World’s First General-Purpose Intelligent Agent “Tongtong” 3.0 Debuts at Zhongguancun Forum — Embodied AI National Standard Released Simultaneously】

【Content】 On March 29, the Beijing Institute for General Artificial Intelligence unveiled “Tongtong” 3.0, the world’s first general-purpose intelligent human agent, at the 2026 Zhongguancun Forum Annual Conference. Its core engine “TongNao” was also revealed and selected as a major forum achievement. Tongtong 3.0 achieves systematic breakthroughs across three dimensions: spatial intelligence (distinguishing 3D virtual from 2D real-world environments); cognitive intelligence (possessing cognitive abilities equivalent to a 5–6 year old child, with emotional interaction capability); and behavioral intelligence (autonomous task planning and physical robot command). On the same day, China’s Ministry of Industry and Information Technology issued the industry’s first embodied intelligence standard (YD/T 6770–2026), effective June 1, 2026, providing a unified framework for standardized industry deployment.

Key Takeaway: A milestone achievement in Chinese AGI research. The national standard’s rollout will accelerate commercialization of industrial and service robots, driving a trillion-yuan embodied intelligence ecosystem.

🔗 Science & Technology Daily

【iQIYI Launches “NaDou Pro” — China’s First Professional Film & TV Production AI Agent】

【Content】 On March 30, iQIYI announced that its AI agent platform “NaDou Pro” has officially opened to the public, entering pre-commercial operation. NaDou Pro is purpose-built for long-form film and television content creation, supporting an end-to-end workflow from script generation and storyboard design to final output. The platform features proprietary “material broad-search” and “blockbuster prompt” capabilities. It integrates iQIYI’s self-developed Qizhi large model alongside multiple mainstream video models including Jimeng, Kling, Vidu, Hailuo, and Wan. Liu Wenfeng, iQIYI Senior Vice President, stated: “The boundaries of film and television creation are being redefined by AI.”

Key Takeaway: A milestone in AI-driven film and TV production, signaling an impending reshaping of content industry productivity patterns. The full deployment of AI + film/TV content toolchains will also drive significant cost compression in digital marketing creative production.

🔗 MSN/Sina Tech

3. AI Investment & Startups

【Eli Lilly Bets $2.75 Billion on Insilico Medicine — Is This Pharma’s “GPT Moment”?】

【Content】 On March 29 (local time), U.S. pharmaceutical giant Eli Lilly and Hong Kong-listed AI drug discovery company Insilico Medicine (03696.HK) announced a licensing and global R&D collaboration valued at up to $2.75 billion. The agreement includes $115 million upfront payment and over $2.6 billion in milestone payments. The partnership leverages Insilico’s end-to-end AI drug discovery platform Pharma.AI to accelerate novel therapy discovery and development across multiple therapeutic areas. Insilico’s generative AI can directly “design” entirely new molecular structures, potentially compressing the traditional drug discovery timeline from over 10 years to under 18 months.

Key Takeaway: The largest single AI pharma investment by a traditional pharmaceutical giant signals that AI drug discovery is moving beyond “storytelling” into genuine commercialization. The biotech-medical AI sector warrants sustained attention.

🔗 36Kr

【All 11 xAI Co-Founders Have Now Left — Musk Completes Total Restructuring】

【Content】 According to Business Insider, with co-founder Ross Nordeen’s departure in late March, xAI—founded by Musk in 2023 with an 11-person team—has now seen its entire founding team exit. Since January 2026, eight co-founders have departed successively, with multiple core project teams facing reductions. The exodus concentrated around the period of the xAI-SpaceX merger restructuring. Musk stated that xAI’s “initial architecture was wrong” and he is rebuilding from the ground up. At present, only Musk himself remains as core leadership, though Grok product iterations continue.

Key Takeaway: A governance risk warning for AI startups. Whether xAI can maintain competitiveness following complete personnel turnover ranks as the most noteworthy organizational experiment in the 2026 large-model race.

🔗 IT Home

【NVIDIA Forward P/E Hits 7-Year Low — AI Hardware Sector Enters Valuation Repricing Phase】

【Content】 A triple confluence of factors—geopolitical tensions (Iran war concerns), inflationary pressure, and market skepticism about AI capital expenditure sustainability—has put sustained downward pressure on NVIDIA’s stock, pushing it near technical bear-market territory (down approximately 20% from recent highs). Forward P/E has fallen to roughly 19.6x, the lowest level since early 2019, below the overall S&P 500 average valuation. Institutional analysis indicates current prices largely reflect pessimistic expectations of “significantly slowing AI spending,” although downside risks have not been fully cleared.

Key Takeaway: NVIDIA’s 7-year low P/E serves as a key barometer of AI infrastructure investment sentiment. While differentiation is emerging in the AI application layer, quality computing infrastructure remains a long-term core theme despite elevated short-term volatility risks.

🔗 Sina Finance

4. Digital Marketing Updates

【Meta Advertising Strategy Fully Upgraded: AI-Generated Video + Creator Collaboration Cuts CPA 19%】

【Content】 Meta recently announced a series of advertising solution upgrades at Shoptalk and IAB NewFronts: ① testing AI-generated UGC-style videos, AI dubbing, and auto-translation tools to improve ad click-through and conversion rates; ② expanding “Trending Reels Ads” coverage to cultural moments including Fashion Week and F1, launching a 24-hour high-attention ad purchase window; ③ upgrading Instagram Creator Marketplace and the “Partner Ad Center,” with data showing creator-partnership ads reducing average customer acquisition costs by 19%; ④ integrated video catalog ad campaigns delivering 20% cost-per-conversion improvement.

Key Takeaway: Meta’s advertising AI acceleration is intensifying. Brands should proactively prepare for AI material production and creator ecosystem partnerships to adapt to systematic transformation in advertising deployment approaches.

🔗 ChuHaiWang

5. Frontiers Research & Industry Insights

【Hangzhou Court Rules AI Provider Liable for Inaccurate Information — Landmark AI Legal Liability Case】

【Content】 The Hangzhou Internet Court concluded China’s first tort dispute case arising from generative AI providing inaccurate information, ruling that the relevant AI platform developer bears civil liability. The judgment clarifies the boundary of developer duty-of-care obligations for AI-generated content (AIGC): if information provided by AI causes verifiable substantive harm to users, developers must bear responsibility. This case has triggered widespread industry questioning about the effectiveness of AI platform “disclaimer” language.

Key Takeaway: The first landmark judicial precedent on AI legal liability. It will compel large model platforms to strengthen output quality control and disclaimer mechanism design, carrying milestone significance for enterprise AI deployment compliance requirements.

🔗 Sina AI Trends

Sources: Quantum Bit, CLS, 36Kr, IT Home, CNMO Tech, Science & Technology Daily, Sina AI Trends, ChuHaiWang, MSN Tech

ZanSpark Generated: 2026-03-31 08:35

This briefing aggregates publicly available RSS feeds for AI-generated content annotation, source attribution, and non-commercial use only. Please contact us for removal requests.

For more AI news, special reports, and blogs, visit 🔗ZanSpark

Today’s Key Trends at a Glance