Sugon announced the completion of the fully domestic AI super cluster 「Sugon 8000 (Dengfeng)」, featuring 100,000 domestically produced accelerator cards and a native supercomputing-intelligent computing fusion architecture. A single system can handle full-precision tasks ranging from FP64 scientific computing to INT8 large model training. The entire technology chain is self-developed, including Hygon chips, scaleFabric high-speed networking, ParaStor storage (ranked first in both IO500 lists), and immersion phase-change liquid cooling. Real-world tests have already used 80,000 cards for protein folding acceleration, 88,000 cards for turbulence simulation, and 90,000 cards to simulate 3.16 trillion atoms. The system supports multi-brand accelerators and mainstream ecosystems, connects to the National Supercomputing Internet, and a second system has begun construction in Zhengzhou to support nationwide computing resource scheduling.

Google has open-sourced Gemma 4, which adopts an encoder-free unified architecture by removing traditional visual and audio encoders and retaining only a minimal projection layer. This allows raw image patches and audio waveforms to be directly mapped into a unified Transformer space, enabling end-to-end 「see, hear, think」 processing. Through the <|think|> control token, all on-device models including 2.3B and 4.5B variants can enable slow-thinking deep reasoning. Released under the Apache 2.0 license, it allows full modification. Technical reports show that Gemma 4 31B Thinking outperforms its larger predecessor 27B model in complex mathematics, frontier science, and Agent tool usage, and directly challenges larger closed-source models on certain high-difficulty benchmarks.

Ant Group’s Robbyant has launched LingBot-VA 2.0, a causal video-action foundation model for embodied robotics. Instead of adapting video generators, it pre-trains a causal DiT from scratch, avoiding the mismatch between bidirectional video objectives and forward control. The semantic vision-action tokenizer builds upon RepWAM with added semantic alignment and latent action targets, allowing unlabeled web videos to provide action-relevant supervision. The video backbone uses a sparse MoE with 128 SwiGLU experts and top-8 routing, activating approximately 2.5B parameters per token out of a total 15.3B. Optimizations including consistency distillation, FP8 TensorRT engine, and paged KV cache reduce single-chunk inference latency from 927ms to 142ms, increasing asynchronous control frequency from 35Hz to 225Hz.

OpenAI's GPT-5.6 has been selected by Microsoft as the preferred model for 365 Copilot, fully integrated into productivity suites such as Word, Excel, and PowerPoint, becoming the default for hundreds of millions of users. On the DeepSWE coding benchmark, GPT-5.6 Sol achieves a 73% success rate at approximately $8.4 per task, significantly outperforming Claude Fable 5 (about 70% success rate at $21.6 per task). The mid-tier GPT-5.6 Terra reaches 70% success at just $4.95. This move is seen as a public response to Microsoft's attempts to replace OpenAI with its own MAI models, marking a shift in office AI competition from 「who is stronger」 to a cost-efficiency race focused on 「value per token」.

Zhipu AI founder Tang Jie released an internal memo announcing that the company will concentrate resources over the next two years on four key technical directions: long-horizon tasks, autonomous agents, self-evolution, and safety governance, temporarily pausing short-term commercialization to push the limits of next-generation model capabilities. Citing METR data, the memo notes that the duration of expert-level human tasks achievable by models doubles every 90 days, evolving from 2 seconds with GPT-2 to over 5 hours with current flagship models, with expectations of models capable of multi-week tasks emerging in two to three years. Zhipu will elevate mechanistic interpretability to a core strategic priority, investing billions to study neuron-level decision logic, aligning with Anthropic’s deep commitment to interpretability.

MTRI released the 「ICML 2026 Panorama Report」, analyzing 6,341 papers via eight LLM agents to reveal the Sino-U.S. academic competition landscape. China leads with 2,446 first-author papers compared to the U.S.'s 1,675, but the U.S. shows higher Spotlight acceptance rates (10.3% vs. 6.3%), reflecting divergent paths between quantity and quality. Tsinghua University ranks first globally with 240 papers, with seven of the top eight institutions being Chinese universities. Industry-academia collaboration accounts for 31.2% of papers. 2026 marks the 「Year of the Agent」, as LLM Agents emerge as an independent research field with 399 papers, surpassing NLP subdomains. 「RL for Reasoning and Post-Training」 becomes the largest sub-theme with 189 papers, signaling a shift in AI research focus from parameter scaling to reasoning and tool use.

360’s Zhou Hongyi analyzed how Anthropic’s Mythos model automates AI-driven vulnerability discovery and introduced 360’s efforts to build a Chinese counterpart using 「Tulongfeng」 and 「Yizhen Tianzhen」. He emphasized that Mythos’ power lies not in the base model alone—which is unstable—but in the agent framework (harness) that stabilizes and unlocks value, meaning capability gaps in base models can be compensated by harness design. Automated vulnerability discovery enables a 「second unilateral transparency」, where attackers can scan in parallel while defenders still rely on scarce experts. 「Tulongfeng」 has identified 3,432 vulnerabilities (including 105 critical ones), while 「Yizhen Tianzhen」 enables autonomous planning and response in real network environments. He highlighted agent unpredictability as the primary security risk, requiring mitigation through workflow structuring and privilege isolation.

A LessWrong study reveals overthinking in reasoning models: answers are typically determined by around 30% of the Chain-of-Thought (CoT), yet models continue 70% more reasoning. By truncating trajectories at various points and forcing early answers, researchers found the 「sufficiency point」 occurs far earlier than actual stopping. They identified a 「termination circuit」—a small group of late MLP layers (primarily layer 27)—that acts as a verification gate, triggering </think> only when the stated answer matches internal computation, rather than when reasoning is sufficient. This gate operates via high-dimensional patterns, not simple directional signals, making it difficult to steer. The most direct way to reduce overthinking is to prompt models to state candidate answers earlier.

On July 10, 2026, Apple filed a lawsuit against OpenAI, accusing it of stealing trade secrets to develop AI devices. On the same day, Google Research unveiled SensorFM, a wearable health foundation model pre-trained on one trillion minutes of sensor data, aimed at enhancing health monitoring. A Northwestern University research team proposed a new AI architecture inspired by the cerebellum to improve coordination and reaction capabilities. Meanwhile, Meta removed Instagram’s AI image generator following public backlash.

AI drama pioneer Jiangyou stated the industry has shifted from 「efficiency dividend」 to 「slow death」. Seedance 2.0 reduced production costs per episode from 150,000 to 40,000–50,000 yuan, with monthly output surging to 180,000 episodes, far exceeding user consumption capacity. The market has devolved from content competition into a game of probability. Regulatory scrutiny and platform throttling accelerate decline—authorities worry about AI dramas disrupting live-action drama supply chains, leading platforms like Douyin and Hongguo to suppress AI content visibility. Going global is the only viable path, but the core barrier is user acquisition scale, not content quality, making it hard for small and medium creators to match platform giants. He described ByteDance’s strategy as 「the last copper coin」, asserting that tool-based companies are inevitably squeezed by large tech firms.