Keywords：Gemini 3 Flash, GPT-5.2, AI model, humanoid robot, Tencent large model, SAM Audio, MiMo-V2-Flash, multimodal reasoning, AI audio editing, embodied robot production capacity, open-source large model, AI health large model
🔥 Spotlight
Google Releases Gemini 3 Flash: Google has launched Gemini 3 Flash, offering cutting-edge intelligence at an ultra-low cost—priced at just 25% of the Pro version. Despite its affordability, it outperforms Gemini 3 Pro on key benchmarks like MMMU-Pro and SWE-bench Verified, with speeds up to 3 times faster. Designed for iterative development, it delivers programming performance and low latency close to Gemini 3 Pro, supporting multimodal reasoning for tasks like video analysis, data extraction, and visual Q&A. Gemini 3 Flash is now widely available: general users can access it via the Gemini app and Google Search AI mode, while developers can integrate it via API, paving the way for broader AI adoption in everyday apps and business systems. (Source: 36Kr, 36Kr)
OpenAI Unveils GPT-5.2 Series: OpenAI has introduced the GPT-5.2 series, including Pro, Thinking, and Instant models, to compete with Google’s Gemini 3. GPT-5.2 sets new records on abstract visual puzzle benchmarks like ARC-AGI-1 and ARC-AGI-2, with significantly improved computational efficiency—costs reduced to 1/390 of GPT-5.1. The models feature adjustable reasoning levels (including x-high) and support context summarization to expand input limits. This release aims to enhance efficiency in professional tasks like code generation, spreadsheet work, and presentations, emphasizing its economic viability for complex reasoning tasks. (Source: DeepLearning.AI Blog)
AI Model Solves Open Math Problem Autonomously: GPT-5 has autonomously solved an open math problem submitted to the IMProofBench benchmark, providing a complete and correct proof without human prompts or intervention. This breakthrough, though small, represents a novel contribution to enumerative geometry and highlights AI’s growing capability in complex mathematical reasoning. The achievement has sparked discussions about AI’s potential in mathematical research, signaling its future role in scientific discovery. (Source: SebastienBubeck, kevinweil, gdb, OfirPress, SebastienBubeck)
Meta Launches SAM Audio: Meta has introduced SAM Audio, the first unified AI model that lets users isolate and edit sounds from complex audio mixes. Like Photoshop for audio, it uses text, visual, and temporal prompts to extract guitar tracks from songs, filter traffic noise, or remove dog barks from podcasts. This innovation could revolutionize audio editing, making it as intuitive as text or image processing, lowering the barrier for professional mixing, and boosting audio content creation. (Source: ylecun, ylecun, dotey)
🎯 Trends
Yunpeng Tech Releases AI+Health Products: Yunpeng Tech, partnering with Shuai Kang and Skyworth, has launched AI+health innovations, including a “Digital Future Kitchen Lab” and smart refrigerators equipped with an AI health model. The AI health model optimizes kitchen design and operations, while the fridge’s “Health Assistant Xiao Yun” offers personalized health management. This marks AI’s deeper integration into daily health and smart home ecosystems, promising tailored wellness services via smart devices. (Source: 36Kr)
Humanoid Robot Industry Faces Production Bottlenecks: In 2025, the embodied robot industry has cooled after initial hype, with orders exceeding ¥3.5 billion but deliveries under 1,000 units, exposing production challenges. Leaders like Zhiyuan and Ubtech struggle to scale from hundreds to thousands of units. C-end adoption remains sluggish due to high prices and limited use cases, as seen in robots’ unstable marathon performances. Despite policy and capital support, the industry is transitioning from “technically feasible” to “reliable commercial products.” (Source: 36Kr, 36Kr)
Meitu’s AI Pivot and Challenges: Meitu refocuses on tools, leveraging apps like MeituPic and BeautyCam for C-end users, and Meitu Design Studio/KaiPai for B-end productivity. Its AI Agent platform, RoboNeo, combines third-party and in-house Miracle Vision models to address AI adoption pain points via “AI generation + manual tweaks.” However, RoboNeo’s user retention lags, especially amid competition like Google’s Nano Banana, raising uncertainties in overseas B-end growth. With 200M MAU and innovation-driven teams, Meitu aims to carve a niche in vertical SaaS. (Source: 36Kr)
Xiaomi Open-Sources MiMo-V2-Flash: Xiaomi has open-sourced its 309B-parameter MiMo-V2-Flash model, joining the top tier of open-source models with 2.6x faster inference while balancing performance and deployment costs. It scores 73.4% on SWE-Bench Verified, surpassing DeepSeek-V3.2 and Kimi-K2 Thinking, and excels in Agent tasks. Innovations like 5:1 hybrid attention, learnable attention bias, MTP multi-layer token prediction, and MOPD multi-teacher distillation boost efficiency and long-text comprehension. Its real-world coding prowess and “high-EQ” interactions reflect Xiaomi’s ambition for “physical-world AGI.” (Source: 36Kr)
Tencent Restructures AI R&D: Tencent has reorganized its AI R&D, appointing former OpenAI researcher Yao Shunyu as Chief AI Scientist under the CEO/President’s Office, reporting directly to President Martin Lau. Yao also heads the AI Infra and LLM departments, signaling Tencent’s push to consolidate competitiveness in compute, data, and engineering. This elevates AI to a core strategy, addressing lagging model development against rivals like ByteDance and Alibaba. (Source: 36Kr, 36Kr, 36Kr)