腾讯研究院AI速递 20251223

Group 1: Generative AI Developments - Gemini 3 Flash outperformed Gemini Pro with a score of 78% in SWE-Bench Verified tests, surpassing Pro's 76.2%, and is 3 times faster than 2.5 Pro while reducing token consumption by 30% [1] - MiniMax has open-sourced its VTP (Visual Tokenizer Pre-training Framework), discovering a Scaling Law in AI visual generation, which resolves the paradox of training performance [3] - Tongyi Qwen launched the Qwen-Image-Layered model, which disassembles images into multiple RGBA layers for independent manipulation, enhancing high-fidelity editing capabilities [4] Group 2: Company Updates and Financial Performance - MiniMax is preparing for an IPO in Hong Kong, with a team of 385 people averaging 29 years old and having spent $500 million, which is less than 1% of OpenAI's expenses [5] - MiniMax reported revenue of $53.44 million for the first nine months of 2025, a year-on-year increase of over 170%, with over 70% of revenue coming from overseas [6] Group 3: Technological Innovations - Shanghai Jiao Tong University introduced the LightGen chip, expanding photonic computing into large model semantic media generation, achieving high-resolution image generation and outperforming NVIDIA's A100 by two orders of magnitude [7] - DeepMind's research suggests that AGI may emerge from multiple smaller AGI agents collaborating rather than from a single large model, proposing a four-layer defense framework for distributed risks [8]