The article reviews the proliferation of new large models in China's tech scene since spring 2023, marking the start of the 'Hundred-Model War'. It emphasizes that these models are not all the same, each with its own characteristics.
Sina Finance reports that a nationwide domestic computing power large model has been officially released, with Southern Fund's Wu Chunlin focusing on domestic tech core assets. This event marks progress in China's independent computing power large model field, potentially impacting investment layouts in the domestic AI industry chain.
IT Home reported on September 7 that an AI enthusiast had OpenAI's new model GPT-6 Astra autonomously complete the 3D puzzle game Portal, involving 3,336 tool calls costing $571.18 at API prices, but actually covered by a $200/month Codex Pro subscription. The experimenter CozyBlaze said the run should not be seen as a standardized benchmark of AI capability, but it demonstrates the potential of general-purpose agents.
iFlytek released the Spark X2.5 large model with 293 billion parameters, trained entirely on domestic computing power. The article includes a hands-on test of the model, showcasing its performance.
Unitree released a video claiming that its UnifoLM-X2-1.0 model has broken through bottlenecks in world-action large model instantaneous planning, decision-making, and dynamic interactive execution, achieving high dynamics, strong interaction, real-time prediction and planning, thus realizing fully autonomous humanoid robot fighting for the first time. This validates the basic feasibility of large-scale deployment of humanoid robots driven by world models.
Unitree announced the world's first real-time world model-driven fully autonomous humanoid robot fighting. Its UnifoLM-X2-1.0 model breaks through bottlenecks in instantaneous planning, decision-making, and dynamic interactive execution, achieving high-dynamic, strong-interaction real-time prediction and planning. This validates the basic feasibility of large-scale deployment of world model-driven humanoid robots.
JD's JoyAI App has launched the first AI digital human capable of real-time video shopping, allowing users to browse and purchase within a video call without switching apps. The digital human supports customization of appearance and persona, and integrates lifestyle services such as food delivery and health consultations. Since the beginning of this year, the number of conversational users on JoyAI App has grown more than 15-fold.
Faster-than-playback generation from models such as MiniMax H3 Max is shifting AI video from batch production to continuous, interactive media. Early experiments show chat-driven infinite livestreams and near-zero-latency narrative games, though consistency, cost and safety remain open constraints.
The article discusses how PC makers like AMD are leveraging Chinese AI models such as DeepSeek, Qwen, and Zhipu for on-device deployment to reduce the high cost of AI tokens, thereby promoting the adoption of AI PCs. This marks a new phase of collaboration between AI models and hardware vendors, aiming to make AI applications more affordable and widespread.
IT Home reports on Arena platform's 36th week of 2026 AI model blind test rankings, showing multiple Chinese models breaking through in specific categories, such as Alibaba's qwen3.8-max-0902 debuting at 4th place in the frontend development list, and Alibaba's wan3.0 entering the top three in the video list. The overall leaderboard is still dominated by overseas models, but the progress of Chinese models in niche areas is noteworthy.
The article reports Nvidia CEO Jensen Huang declaring that AGI has arrived, and discusses the computing power demands behind the GPT-6 Astra model, indicating that related sectors will see growth. It emphasizes the critical role of computing power in AI development and implies positive implications for hardware vendors like Nvidia.
OpenAI has made initial progress in automated research, signaling an accelerating trend of 'AI training AI.' The article notes that despite AI's growing capabilities, compute remains a critical bottleneck for future development.
The article discusses OpenAI's demonstration of 'alien thinking' after the release of GPT-6, sparking debate on whether AGI has truly arrived. It likely covers model capability advancements and AGI progress, but provides no specific financial data or investment information.
The article discusses the effectiveness of next-chunk reasoning reinforcement learning in AI model training, noting that despite 60x more training compute, model performance worsened, questioning the viability of this technical approach.
This week, tech giants including Anthropic, Meta, Google, and OpenAI released rapid updates to AI models, sparking concerns about 'model fatigue.' OpenAI CEO Sam Altman said everyone is accelerating iteration. Runpod CEO noted the current environment has too much froth, forcing companies to generate noise. Meanwhile, Anthropic and OpenAI are each valued near $1 trillion in private markets, and global AI spending is projected to reach $2.59 trillion this year, up 47%.
This week, Anthropic, OpenAI, Meta, and Google all released model updates, while Nvidia announced it is acquiring open-source AI platform Hugging Face. The frenetic release pace is causing 'model fatigue,' suggesting market excitement over AI innovation may be waning.
Microsoft is adjusting its Windows 11 AI strategy, shifting from emphasizing Copilot buttons to the concept of 'unmetered intelligence,' which involves running more AI models and agents directly on local PCs, reducing reliance on the cloud. This reflects the industry trend from chatbots to agents and may reduce users' need for subscription services.
Alibaba Qwen open-sourced Qwen-Drive-1.0-4B this week, a vision-language model for autonomous driving built on Qwen3.5-4B, claimed to be the first vision-language foundation model for autonomous driving. The model unifies 3D perception and visual question answering during pretraining, extends to motion planning, while retaining general vision-language capabilities. It offers both SFT and RL planning experts, with official evaluations showing competitiveness across benchmarks and improved closed-loop safety after reinforcement learning.
OpenAI's head of applied research, Boris Power, stated that the Astra model has surpassed humans in spatial reasoning, closing one of humanity's last advantages. Independent researcher tests show Astra performs flawlessly, solving previous LLM vision issues. OpenAI calls Astra the 'most intelligent and aligned model' and declares 'Welcome to the AGI era.'
Tesla has started rolling out FSD v14.3.9 Supervised, adding an active safety mechanism that can engage FSD on behalf of the driver during manual driving if an imminent collision is detected and automatic emergency braking is insufficient, executing coordinated evasive maneuvers involving steering, braking, and acceleration. The feature is positioned as a last-resort safety net, not a replacement for driver attention, and only works when FSD is active.