Wall-facing intelligence

On September 19th, yesterday afternoon, the smart face wall announced a new series of "Small Steel Guns": the VoxCPM base model for voice generation of 0.5B parameter sizes was introduced. The VoxCPM base model for voice generation was officially launched at the Human Voice Interactive Laboratory of the International Graduate School of Shenzhen, Singhua University. The model parameter size is 0.5B, with an industry SOTA level in terms of voice naturality, sound similarity and rhythm performance. Performance performance: RTF ≈ 0.17, support current output VoxCPM in Seed-TTS-EV..

August 27 news, the face of the wall intelligent August 26 announced open source 8B parameters of the face of the wall MiniCPM-V 4.5 multimodal flagship model, becoming the industry's first multimodal model with "high brush" video comprehension capabilities. MiniCPM-V 4.5 is said to have high brush video comprehension, long video comprehension, OCR, document parsing capabilities of the same level of SOTA, and the performance exceeds the Qwen2.5-VL 72B, so it is called "the strongest end-side multimodal model". Facade Intelligence said that the previous mainstream multimodal model in dealing with video comprehension tasks,...

According to the OpenBMB open source community, MiniCPM-V 4.0, a new generation of multi-modal model of the face wall intelligent small steel gun, was officially open sourced yesterday. It is reported that MiniCPM-V 4.0 relies on 4B parameters to achieve the same level of SOTA results in OpenCompass, OCRBench, MathVista and other lists, and realize a stable, silky smooth operation on cell phones. Specifically, MiniCPM-V 4.0 achieves the same level of SOTA results on OpenCompass, OCRBench, MathVista and other lists, and realizes stable and silky smooth operation on mobile phones.

June 7 news, the face of the wall of intelligence on the evening of June 6 released end-side large model MiniCPM 4.0, the company said the new model through self-research CPM.cu reasoning framework, in extreme scenarios to achieve the highest 220 times speedup, the regular 5 times speedup, support in vLLM, SGLang, LlamaFactory and other frameworks deployed. The released 8B Lightning Sparse Edition adopts the innovative sparse architecture to set off a high efficiency storm, while the other 0.5B is called "the most powerful small steel gun that is lightweight and agile". According to the official introduction, the...

Jan. 16, 2012 - Facade Intelligence announced today the launch of MiniCPM-o 2.6 end-side full-modal model with 8B parameters, claiming performance comparable to that of GPT-4o and Claude-3.5-Sonnet. It adopts an end-to-end multimodal architecture, and can process multiple types of data, such as text, images, audio and video, to produce high-quality text and speech outputs simultaneously. , audio and video and other types of data simultaneously, generating high-quality text and speech output. Officially, its total number of parameters is 8B, and its visual, speech, and multimodal streaming capabilities have reached the GPT-4o-202405 level, making it one of the most popular open source communities for modal support....

The official WeChat account of Mianbi Intelligence published a blog post yesterday (September 5), announcing the launch of the open source MiniCPM3-4B AI model, claiming that "the end-side ChatGPT moment has arrived." This is an excellent AI model that can run on devices with only 2GB of memory, heralding a new era of end-side AI experience. The MiniCPM3.0 model has 4B parameters and outperforms GPT-3.5 in performance, and can achieve AI services on mobile devices at the same level as GPT-3.5. This allows users to enjoy fast, secure, and functional...

Mianbi Intelligence and Huawei Cloud reached a formal cooperation at the Huawei Developer Conference 2024 (HDC2024) on June 22. Mianbi Intelligence is the first end-side big model partner in Huawei Cloud's promotion of big model end-cloud collaborative solutions. The two parties will jointly promote the development and deployment of big model end-cloud collaborative solutions to further reduce the cost of using big models and improve response speed; put big models "closest to users" and further expand the application value of big models in various scenarios and different fields. Mianbi's core R&D team was born out of Tsinghua Natural Language Processing Laboratory. Previously, their open source...

Recently, the Llama3-V open source model of the AI research team at Stanford University was accused of plagiarizing the open source model "Mini Cannon" MiniCPM-Llama3-V 2.5 developed by the star startup company Mianbi Intelligence of Tsinghua University, which caused heated discussions online. Image source Pexels On May 29, a Stanford AI team claimed online that it only needed $500 to train a SOTA multimodal large model that surpassed GPT-4V, but netizens soon discovered that the model structure and code used in the project were highly similar to the "Mini Cannon", with only...

Mianbi Intelligence released the open source large model Eurux-8x22B, including Eurux-8x22B-NCA and Eurux-8x22B-KTO, focusing on reasoning ability. In the official test, Eurux-8x22B surpassed Llama3-70B in LeetCode (180 LeetCode programming questions) and TheoremQA (IT Home Note: STEM questions at the level of American universities) tests, and surpassed the closed source GPT-3.5-T in LeetCode tests...

The latest flagship end-side model launched by Mianbi Intelligent - Mianbi MiniCPM2.0 series models bring a series of amazing performance and functions: 1. MiniCPM-V2.0 is the strongest multimodal model on the end-side, with powerful OCR capabilities, and even some capabilities are comparable to Gemini Pro. It can accurately recognize various complex image contents, including street scenes and long pictures, through its self-developed high-definition image decoding technology. 2. MiniCPM-1.2B is a base model that is more suitable for end-side scenarios, and its performance exceeds many mainstream models,…

Mianbi Intelligence announced yesterday that it had completed a new round of financing of several hundred million yuan, led by Primavera Ventures and Huawei Hubble, followed by Beijing Artificial Intelligence Industry Investment Fund and others, and Zhihu continued to follow up as a strategic shareholder. Hubble Technology Venture Capital Co., Ltd. is wholly owned by Huawei Investment Holding Co., Ltd. and was established on April 23, 2019. The legal representative, chairman and general manager is Bai Yi, president of Huawei Global Financial Risk Control Center. Previously, Huawei Hubble invested more in semiconductor companies. This is the first time Huawei Hubble has invested in a domestic AI large model company. Mianbi's core R&D team was born out of Tsinghua Natural Language…

Mianbi Intelligence recently completed a new round of financing of several hundred million yuan, led by Primavera Ventures and Huawei Hubble, followed by Beijing Artificial Intelligence Industry Investment Fund and others, and Zhihu continued to follow up as a strategic shareholder. This round of financing will be used to promote the introduction of outstanding talents, strengthen the underlying computing power and data foundation for the development of large models, continue to lead the "efficient large model" route, and promote the efficient training and rapid application of large models. Mianbi Intelligence is one of the largest model teams in the world that is at the forefront of exploring "efficient large models". The core R&D team comes from Tsinghua NLP Laboratory, and the team members include well-known companies such as Alibaba, ByteDance, and Baidu...