Today's Overview

Domestically, Zhhipu launches GLM-5.1 high-speed API and breaks the global latency record, Baidu's Wenxin team releases PaddleOCR-VL-1.6 document parsing model, Alibaba's Tongyi Qianwen Qwen3.7-Plus surpasses GPT-5.4 in screen understanding benchmark, and Tencent Hunyuan simultaneously unveils Stem sparse attention algorithm and Hy-Memory long-term memory plugin; internationally, Anthropic open-sources AI vulnerability discovery framework and discloses "AI self-improvement" research, Alphabet announces $80 billion equity financing to expand AI computing power, and Liquid AI releases 8B MoE new model LFM2.5, as the industry continues to accelerate along three main lines: "model capabilities + computing infrastructure + Agent engineering".

Domestic Highlights

🔥 Zhhipu Releases GLM-5.1 High-Speed Version, Setting New Global Large Model API Speed Record

Zhhipu released the GLM-5.1 high-speed API on June 4, with measured first-token latency reduced by 60% compared to the previous generation, setting a new global large model API speed record. The GLM Coding Plan pricing tier was launched simultaneously, selling out immediately upon release. On the capital market, Zhhipu's Hong Kong-listed shares surged over 20% in early trading.

Source: Pingwest | Details: https://www.pingwest.com/search/?q=GLM-5.1

🔥 Alibaba's Tongyi Releases Qwen3.7-Plus: Outperforms GPT-5.4 in Screen Understanding, Develops App Independently in 11 Hours

Tongyi Qianwen launches Qwen3.7-Plus, leading GPT-5.4 in screen understanding benchmark, and demonstrates the complete process of an app developed independently by the model within 11 hours, emphasizing end-to-end "seeing, thinking, writing, doing" capabilities to further enter the AI Agent track.

Source: Wall Street CN | Details: https://wallstreetcn.com/search/?keyword=Qwen3.7-Plus

🔥 Baidu ERNIE Bot Releases PaddleOCR-VL-1.6: Document Parsing Accuracy Exceeds 96.33%

Baidu ERNIE Bot team released PaddleOCR-VL-1.6 multimodal document parsing model, setting new records on multiple document understanding SOTA benchmarks with 96.33% accuracy, and has been released for download on PaddlePaddle and ERNIE Bot ecosystem, focusing on enterprise-level OCR/document intelligence scenarios.

Source: QbitAI | Details: https://www.qbitai.com/search/?keywords=PaddleOCR-VL-1.6

🔥 Tencent Hunyuan Proposes Stem Sparse Attention: First-Token Latency Reduced by 3.6x in Long-Context Reasoning

Tencent Hunyuan team proposes Stem sparse attention algorithm, reducing first-token latency by 3.6x in long-context reasoning, achieving new SOTA on long document summarization and code completion tasks. Paper and inference code have been open-sourced, targeting the pain points of large model long-context deployment.

Source: Tencent Hunyuan | Details: https://hunyuan.tencent.com/news

🔥 Tencent Hunyuan Launches Hy-Memory Memory Plugin, Reshaping Long-Term Collaborative AI Experience

Tencent Hunyuan released Hy-Memory long-term memory plugin, enabling AI assistants to沉淀 user preferences and project context across sessions. Official claims can improve 30-day collaborative task success rate to 78%, providing infrastructure for "collaborative AI Agent."

Source: Tencent Hunyuan | Details: https://hunyuan.tencent.com/news

🔥 Domestic Computing Power Completes Full-Parameter Post-Training of Trillion-Parameter AI Large Model

According to Securities Times, domestic GPU clusters have successfully completed full-parameter post-training of trillion-parameter AI large model, marking a key breakthrough in domestic computing power for large model training pipeline, significantly reducing dependence on overseas high-end GPUs.

Source: Securities Times | Details: https://www.stcn.com/search/?keyword=%E5%9B%BD%E4%BA%A7%E7%AE%97%E5%8A%9B+%E4%B8%87%E4%BA%BF

🔥 Doubao Set to Enter Paid Era, ByteDance AI Seeks New Growth Curve

36Kr exclusively revealed ByteDance's four key AI propositions for 2026, with Doubao officially moving towards paid services being an important part of this; at the same time, former Seed head Gu Quanquan departed, and the team restructuring focuses on ToB, model subscriptions, and AI commercialization loop.

Source: 36Kr | Details: https://36kr.com/search/articles/%E8%B1%86%E5%8C%85%20%E6%94%B6%E8%B4%B9

🔥 DeepSeek Completes First Round of Financing of Approximately 500 Billion Yuan, Tops US Enterprise New Procurement List

DeepSeek is about to complete its first round of financing with a scale of approximately 500 billion yuan RMB; simultaneously, it topped the US enterprise new procurement list, becoming a landmark AI product for Chinese companies going global, and an important milestone for domestic LLM commercialization.

Source: Phoenix Technology | Details: https://tech.ifeng.com/search/?keyword=DeepSeek+500%E4%BA%BF

International Hot Topics

🔥 Anthropic Open-Sources AI Vulnerability Discovery Framework, Betting on AI Security Engineering

Anthropic open-sourced a reference harness for AI-driven vulnerability discovery, supporting enterprise security teams to integrate AI vulnerability hunters into CI/CD and red team processes. The repository has been publicly released on GitHub, advancing the engineering of AI security research.

Source: GitHub | Details: https://github.com/anthropics/defending-code-reference-harness

🔥 Anthropic Releases "When AI Builds Itself": Towards Recursive Self-Improvement

Anthropic Institute published a blog post introducing the latest progress in recursive self-improvement, covering model self-debugging, automated evaluation, and training data self-generation, addressing the industry's security concerns about "AI self-evolution."

Source: Anthropic | Details: https://www.anthropic.com/institute/recursive-self-improvement

🔥 Uber Sets $1,500/Month AI Usage Cap: Industry AI Tool Pricing Signal

Simon Willison wrote an article analyzing Uber's implementation of a $1,500/month usage cap on internal AI tools, arguing this is an early signal of enterprise-level AI coding/Agent tool pricing moving toward "hard budgets," also reflecting the reality of rapidly rising token costs.

Source: Simon Willison | Details: https://simonwillison.net/2026/Jun/3/uber-caps-usage/

🔥 Alphabet Announces $80 Billion Equity Financing to Expand AI Computing Power

Alphabet announced plans to raise $80 billion through equity financing, focusing on AI infrastructure, data centers, and computing power expansion, strengthening Google Cloud and Gemini model training and inference capabilities.

Source: ABC News | Details: https://abc.xyz/investor/news/news-details/2026/Alphabet-Announces-Proposed-80-Billion-Equity-Capita

🔥 Liquid AI Releases LFM2.5-8B-A1B: A MoE Model Trained on 38T Tokens

Liquid AI released the LFM2.5-8B-A1B mixture-of-experts model, with 8B total parameters and 1B active parameters, but trained on a massive 38T tokens of data, focusing on low-latency edge inference and high-quality code generation, open source available for download.

Source: Liquid AI | Details: https://www.liquid.ai/blog/lfm2-5-8b-a1b

🔥 Stanford Law School Research: AI Beats Law Professors on Legal Tasks

The latest research from Stanford Law School shows that mainstream large models have stably surpassed senior law professors in tasks such as contract review and legal statute retrieval, sparking widespread discussion in legal education and the legal services industry, and reigniting the "AI replacing white-collar workers" debate.

Source: Stanford Law | Details: https://law.stanford.edu/press/ai-outperforms-law-professors-in-stanford-law-study/

🔥 Anthropic Surpasses OpenAI to Become the World’s Most Valuable AI Startup

According to multiple media reports, Anthropic’s valuation in a new round of financing has surpassed OpenAI, making it the world’s most valuable AI startup, with a valuation ranging around 300 billion USD, intensifying the Matthew effect among leading large‑model companies.

Source: Qazinform | Details: https://qazinform.com/news/anthropic-surpasses-openai-to-become-worlds-most-valuable-ai-startup

---

More Reading