Apple has officially enabled 'Apple Intelligence and Siri' for mainland China users via the iOS 27.3 developer beta on iPhone 15 Pro and newer models. The update integrates Alibaba's Qwen LLM directly into the system, supporting features like real-time translation, writing tools, and visual intelligence without requiring separate app launches.
Multimodal Understanding
多模态理解
The ability of a model to process and correlate multiple data types like text and images simultaneously.
Curated Intelligence
Apple announced the launch of Apple Intelligence in China, integrating Alibaba's Qwen model into Siri, writing tools, and visual intelligence. The feature will be officially released with iOS 27.3 in December, allowing users to opt-in for privacy protection.
Xiaomi has retired the MiMo-V2.5-Pro model, automatically routing requests to the new MiMo-V2.6-Pro which supports native multimodal understanding. While maintaining API compatibility and low pricing, the update increases default context limits but restricts certain sampling parameters in thinking mode.
Google officially released its flagship model, Gemini 4 Argon, demonstrating significant improvements in multimodal capabilities, coding, and complex task execution. Third-party benchmarks place its intelligence index on par with GPT-6 Astra and Claude Opus 5.5. The model emphasizes cost-efficiency, offering API pricing approximately 39% lower than competitors, alongside a limited-time 50% discount to regain market share.
NVIDIA has released Nemotron 3.5 Super VL, an open-source multimodal model optimized for long-running agents that integrates text, image, document, and video understanding. Built on a 120-billion-parameter Mixture-of-Experts (MoE) architecture with only 12 billion active parameters, it balances accuracy and efficiency while supporting custom post-training via NVIDIA NeMo. vLLM provides Day 0 support, enabling immediate deployment for workflows requiring high-precision visual reasoning and tool use.
Apple's filing for the China-specific iPhone Duo confirms the integration of Alibaba's Qwen large model into Apple Intelligence. This partnership enables native text and image generation across iOS, macOS, and visionOS without app switching, coinciding with the foldable device's October 23 launch to accelerate on-device LLM adoption in premium hardware.
Google DeepMind introduces Gemini Robotics 2, a general intelligence layer enabling whole-body control, fine dexterity, and multi-robot collaboration. It moves beyond pre-programmed or teleoperated constraints by using multimodal understanding to drive complex task execution and facilitate skill transfer across diverse robot bodies.
The vLLM-Omni project details its multi-stage serving architecture for the Qwen3-Omni multimodal model, utilizing a pipeline of Thinker, Talker, and Code2Wav stages. By implementing stage-level batching, CUDA Graphs, and asynchronous chunk handoffs, the system mitigates decoding stalls in high-concurrency scenarios and significantly reduces Time to First Audio Packet (TTFP). This work provides a reusable performance optimization paradigm for deploying omni-modality models.
OpenAI officially released GPT-5.4 mini and nano, targeting high-volume workloads with optimal cost-performance ratios. GPT-5.4 mini significantly outperforms its predecessor in coding, reasoning, and multimodal tasks while running over 2x faster, approaching flagship GPT-5.4 performance on benchmarks like SWE-Bench Pro. The nano variant is optimized for low-latency, low-cost tasks such as classification and simple subagent operations.
Google has officially launched Gemini 3 Pro, claiming superior performance in AI benchmarks and coding tasks compared to previous versions. Accessible via Google AI Studio, Vertex AI, and the new Antigravity platform, the model emphasizes multimodal understanding, visual reasoning, and 'vibe coding' capabilities to lower development barriers and support agentic workflows.