AI startup Moondream has officially announced the completion of $4.5 million in seed funding and presents a disruptive viewpoint: in the world of AI models, smaller models may hold advantages. The company is backed by Felicis Ventures, Microsoft's M12 GitHub Fund, and Ascend, launching a visual language model with only 1.6 billion parameters that can compete with models four times its size in terms of performance.

The Allen Institute for Artificial Intelligence (AI2) has released an open-source large language model OLMoE, aimed at providing high performance and low-cost solutions. This model employs a Sparse Mixture of Experts (MoE) architecture with 7 billion parameters, but through an intelligent routing mechanism, only 1 billion parameters are used per input token, achieving efficient computation. OLMoE includes both general and instruction-tuned versions, supporting a context window of 4096 tokens. Its training data is sourced from a wide range, including Common Crawl, Dolma CC, and Wikipedia.

Microsoft has released Phi-3.5-vision, a lightweight, multimodal open source AI model designed for processing textual and visual inputs, supporting a context length of 128K. This model is suitable for resource-constrained environments and features capabilities such as image understanding, OCR, chart parsing, and multi-image summarization, showcasing excellent performance and low latency. Comprised of 4.2 billion parameters, it is trained with high-quality data to ensure performance and privacy. It includes three models: lightweight AI, expert mix, and multimodal model, all demonstrating outstanding performance in image and video processing benchmarks.

Tencent's Hunyuan DiT model has recently partnered with the community to release three new controllable plugins for ControlNet: tile (high-definition image enlargement), inpainting (image restoration and expansion), and lineart (line art generation), further expanding its ControlNet matrix. The addition of these plugins enables the Hunyuan DiT model to cover a wider range of application scenarios.

On July 4, 2024, the Shanghai Artificial Intelligence Laboratory launched the new version of the Shusheng·Puyang series model, InternLM2.5, at the WAIC Scientific Frontier Main Forum. This version has significantly enhanced reasoning capabilities in complex scenarios, supports 1M ultra-long context, and can autonomously perform internet searches and integrate information from hundreds of webpages.

FLUX.1, an open-source large model platform launched by AI image generation expert Robin Rombach, features a Vision Transformer architecture with 12 billion parameters. Utilizing techniques such as flow-matching training methods and rotational positional embedding, it significantly surpasses closed-source models like DALL·E 3 and Midjourney V6, as well as the SD3 series. In terms of performance, FLUX.1 excels in text semantic restoration, image quality, action consistency, coherence, and diversity.

Meta has launched the Llama3.1 open source model series, which includes a 405B parameter version that performs comparably to GPT-4. Among them, the 8B parameter version Llama3.1-8B-Instruct supports multiple languages with a context length of up to 131072 tokens and has been trained on 25 million synthetic data entries, demonstrating cognitive and reasoning capabilities similar to GPT-3.5 Turbo. Based on this model, OpenBuddy has released OpenBuddy-L, which supports Chinese Q&A and cross-language translation.

In the realm of artificial intelligence, the competition between open-source and closed-source models has never ceased. The recent release of Meta AI's Llama 3.1 model, however, seems to have drawn a watershed line in this contest. This is not merely the launch of a new model; it is a sign of the maturation of open-source AI, heralding the arrival of a new era.Llama 3.1, developed by Meta AI's team, is a new generation of large-scale language models. In over 150 benchmark tests, its 405B paramet

DeepSeek recently announced the release of an open-source model called DeepSeek-Coder-V2, which surpasses GPT-4-Turbo in code and mathematical capabilities, and features significant expansion in multilingual support and context processing length. Based on the DeepSeek-V2 model architecture, DeepSeek-Coder-V2 employs a mixture of experts (MoE) architecture specifically designed to enhance code and mathematical reasoning abilities.

In the early hours of today, Alibaba's Tongyi Qianwen team released the Qwen2 series of open-source models. This series includes pre-trained and instruction-tuned models in five sizes: Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B, and Qwen2-72B. Key information indicates that these models show significant improvements in parameter count and performance compared to the previous generation Qwen1.5.

Zero One Everything has announced the open sourcing of the Yi-9B model. Yi-9B is the strongest model in the Yi series in terms of coding and mathematical capabilities. It excels in overall, coding, and mathematical abilities, and can be easily deployed on consumer-grade graphics cards. The company was founded by Kaifu Lee, Chairman and CEO of Innovation Works.

Google has released the open-source large model Gemna, breaking away from last year's closed-source strategy. Google is now engaging in a dual strategy of both closed and open source, competing in the large model market against OpenAI and Meta, forming a trio of giants: Google Gemna, Meta LLama, and Mistral. The competitive landscape in large models is intensifying as Google and Meta vie for dominance.

Google and Hugging Face recently announced a strategic partnership aimed at advancing open AI and machine learning. This collaboration will integrate Hugging Face's platform with Google Cloud's infrastructure, including Vertex AI, to make generative AI more accessible for developers. The partnership supports developers in quickly and cost-effectively training, fine-tuning, and serving open models on Google Cloud, utilizing AI to optimize infrastructure such as TPUs and GPUs.

Mistral AI, established just six months ago, completed two funding rounds of $113 million and $415 million, skyrocketing its valuation from $260 million to $2 billion. Mistral AI released its open-source model Mixtral 8x7B, which performs closely to GPT-4. Mistral AI is regarded as 'France's OpenAI' and has received strong support from President Macron. Mistral AI aims to compete directly with OpenAI through open-source and differentiated positioning.

["Orion Star has released the open source OrionStar-Yi-34B-Chat model based on the Zero One Thousand Things initiative, available for free commercial use.","The model performs excellently on the C-Eval and C-MMLU Chinese evaluation benchmarks.","OrionStar-Yi-34B-Chat is one of the outstanding open source alternatives to ChatGPT.","The model boasts high performance, is fully open-source, and is free for commercial use.","Orion Star is a provider of AI service-oriented solutions."]

Alibaba Group CEO Wu Yongming announced at the 2023 World Internet Conference that Alibaba will open source a 72 billion parameter large model. This will be the largest scale open source model in China. This move signifies Alibaba Cloud's leading position in the field of large models. The open source large model will promote the development and application of artificial intelligence technology. Other companies are also releasing their own open source large models, intensifying competition in AI technology.

1. Yi-34B and Yi-6B are two versions of the Chinese and English large models developed by Dr. Kai-Fu Lee's Zero One Infinity team. 2. The Yi-34B model supports the world's longest 200K context window, capable of processing 400,000 Chinese characters. 3. Yi-34B and Yi-6B performed well in various benchmark tests, ranking highly in MMLU, BBH, and C-Eval. 4. Kai-Fu Lee stated that the Yi series of large models is developing multimodal capabilities, and relevant results are expected to be released in 1-2 months. 5. Kai-Fu Lee believes that large models...

The ModelCache architecture under Ant Group's CodeFuse includes the adapter, embedding, similarity, and data_manager modules. ModelCache can reduce the inference cost of large model applications and enhance user experience. Cache hits can reduce average latency by 10 times, with speed improvements of up to 14.5%. ModelCache will continue to optimize performance and accuracy, improving recall time.

Zhaolian Consumer Finance has launched 'Zhaolian Zhilu', the first open source model with 13 billion parameters in the consumer finance industry, which has received authoritative recognition in the field of general knowledge. This model features high precision and financial expertise, meeting customization needs to improve work efficiency and reduce costs. The model will be applied in subdivided scenarios such as risk control management, code generation, and operational enhancement, while continuing to explore intelligent agent technology and promoting technology for good. Zhaolian Consumer Finance's open source model 'Zhaolian Zhilu' has ranked on authoritative lists C-Eval and CMMLU.