Alibaba has open-sourced its Qwen3-Embedding series, a vector model with 40% performance gains over predecessors, surpassing competitors like Google and OpenAI in text retrieval and classification tasks. The model supports over 100 languages and is now available on multiple platforms.
AI coding assistant Cursor, developed by Anysphere, has secured $900 million in funding, boosting its valuation to $9.9 billion. The company has completed three funding rounds in under a year, with annual recurring revenue doubling every two months and now exceeding $500 million.
Mary Meeker's new report reveals AI training costs are nearing $1 billion while inference costs have plummeted 99%, creating a stark divide that's reshaping the industry's competitive landscape.
A new AI trend report highlights the stark contrast between skyrocketing model training costs, now reaching billions, and rapidly declining inference expenses. The findings suggest this economic shift is fueling a surge in AI experimentation and product development.
NVIDIA has launched the Llama-Nemotron-Nano-VL-8B-V1, an 8B-parameter visual language model supporting image, video, and text inputs. The model excels in tasks like document intelligence and image reasoning while offering efficient deployment on edge devices. Released under an open-source license, it positions NVIDIA as a key player in the multimodal AI space.
Tokyo-based Shisa.AI has released its open-source Shisa V2 405B model, which surpasses GPT-4 in Japanese language processing while maintaining strong English capabilities. The bilingual model represents a significant advancement for Japan's AI sector and is now available to developers worldwide.
Midjourney is set to release its highly anticipated video generation feature this month, alongside significant upgrades to its AI models. The V8 model promises major improvements, while server expansions and style reference optimizations aim to enhance user experience.
Hume AI has unveiled EVI 3, its third-generation voice interaction model that excels in emotional recognition and personalized responses. The AI demonstrates superior speed and emotional intelligence compared to competitors like GPT-4, with applications ranging from customer service to content creation.
Cerebras Systems has fully opened its AI inference API, removing waitlist restrictions and offering developers 1 million free tokens daily. The service boasts speeds up to 20x faster than GPUs, supporting models like Llama4 and Qwen3-32B.
Huawei has unveiled its 'Ascend + Pangu Ultra MoE' system, capable of solving advanced math problems in just two seconds without relying on GPUs. The breakthrough highlights Huawei's advancements in domestic computing power and AI model training efficiency.
New research from Arizona State University reveals that large AI models like ChatGPT don't truly reason but instead identify statistical patterns in data, challenging common perceptions of their cognitive abilities.
Qwen Lab and Peking University unveil ZeroSearch, a revolutionary framework that reduces large language model retrieval training costs by 88% by simulating search engines without real queries.
Meta and The Chinese University of Hong Kong have developed Multi-SpatialMLLM, a revolutionary multimodal AI model that significantly improves spatial reasoning. The model integrates depth perception, visual correspondence, and dynamic perception to overcome limitations in current systems.
Resemble AI has open-sourced Chatterbox, a powerful text-to-speech model that rivals ElevenLabs in performance. With advanced features like voice cloning and emotional control, it's set to transform voice technology applications while maintaining security through neural watermarking.
Lemon AI, a new full-stack intelligent agent, promises to revolutionize task automation across industries with its autonomous capabilities and multi-functional integration.
Meituan's AI advancements include foundational models approaching GPT-4 capabilities, a business decision-making assistant launching in June, and AI-generated code accounting for 52% of internal engineering work. The company is also developing a "NoCode" programming tool for non-technical users.
Alibaba's DAMO Academy has launched the URM General Recall Large Model, a breakthrough in AI-driven advertising that enhances e-commerce ad targeting through deep learning and big data analysis. The technology promises higher ROI for businesses while delivering personalized shopping experiences.
© 2024 - 2025 Summer Origin Tech
