Open Source Model

News from 7 October, local time, 6 October, the French company Mistral AI officially launched the public preview of the Mistral Large 4 model, short ML4, code name le Chonk. As of this date, the user will be able to experience the preview API on the Mistral Studio, and the model weight will be officially available for download at the end of this month. Officially, the top network security agencies, trusted partners and government regulators are now working together in a real business environment..

9 September message, face wall smart release and open source MiniCPM 5-2B. This is a dense model of 2.6 billion parameters, with Artificial Analysis ' s Inteligence Index v4.2 divided into 15 and ranked first in the open source weight model below the 4 billion parameters recorded by the agency. MiniCPM5-2B supports 13.31 million token context, using Apache 2.0 licence, which can be deployed on mobile phones, tablets and edges. Model card..

On August 23, according to the South China Morning Post, a United States artificial intelligence start-up company with OpenAI investments built its first internal model based on Kimi K3 of Moonshot AI Laboratories in China. This highlights the fact that, as AI model development costs are soaring, more and more Western technology companies are moving towards China’s open weight model. Harvey, local time announced on Thursday that the new model of Harvey Tent is open weight..

The goal of this lesson: a computer with only a system and a graphic card, followed by an AI video with stereo, and learned the official hint -- H3 is good and bad, and most of it is on it. At each step, it was written “what to do, how to do, what to see after”, without any code basis. One thing you have to know before you start: the native canvas of the local model is 768p. The model was only trained on the short side of a canvas of 768, and the official 2K came out of another closed-source module. But 768p is not a local ceiling: resolution can be forced..

On July 26, an open letter calling for an open weight AI model was signed by 25 U.S. technology firms this week (24 July). 1AI understands that the open letter refers to the history of open-source software movements and addresses the critical importance of the open weight AI model for the United States to take ownership of the artificial intelligence age. Open weight models, which can be downloaded, modified and deployed autonomously, can lower the threshold for the use of advanced AI by various institutions, compete in dynamic markets, and avoid AI gains being monopolized by a few enterprises, while allowing the use of..

On July 17, Kimi K3 was officially released, using 2.8 trillion parametric scales, from birth to support visual understanding and providing 1 million token context windows. Officially, it is called the world ' s first 3 trillion-billion-scale open source model, which focuses on long-range programming, knowledge work and reasoning. Kimi K3 uses MoE structures to activate 16 of the 896 experts. Officially, Stable Laten Moe, Quantile Balancing, Per-Hea..

On July 9th, an article entitled "Eoone is wrong about open source AI in the environment" was published by the American company Decagon CEO Jesse Zhang, providing an anti-intuitive theory that there is no competitive relationship between open-source models and closed-source models at the enterprise level in the AI market, but rather two different stages in the life cycle of the same product. Zhang takes the example of the company:

In the morning of April 3, Google DeepMind released a new generation of open source model series Gemma 4 with a one-time rollout of four models covering the side to the workstation scene. E2B: 5.1 billion total parameters, 2.3 billion valid parameters, 128K context, officially stated as some equipment memory occupancy up to below 1.5GB; E4B: 8 billion total parameters, 4.5 billion valid parameters, 128K context, MMLLU Pro, 69.4%, close to previous generation 27B level; 26B A4B M..

On February 17, Ali, on the Chat.qwen.ai page, went low on line with two new models, Qwen3.5-Plus and Qwen3.5-397B-A17B. 1AI learned from the official page that Qwen3.5 Plus is located in the latest large language model of the Qwen 3.5 series, and Qwen 3.5-397B-A17B is located in the Qwen3.5 large language model of the Open Source series. Both models support text and multi-mode tasks. According to Ali Yun, Qwen3..

On November 25th, a new open source model, the HunyuanOCR, was announced today, with only 1B, based on a multi-modular structure of the hybrids, with many industry OCR application lists SOTA (note: state of the art). According to official sources, thanks to the conceptual design of the Magnificent Multi-Mode Model “end-to-end” philosophies, the functions of the HunyuanOCR can achieve optimal results only if it is based on a single forward reasoning. The OCR expert model is built on a hybrid multimodular structure and consists of three main components: native..

15 November, at the & OpenEuler Summit General Assembly, Nii Guangnam, Member of the Chinese Engineering College. Ning Gwang-nam stated that the “Fifty-five” plan clearly proposes a high level of openness to the outside world and a new situation of cooperation for win. Our country will pursue a strategy of international scientific and technological cooperation that is more inclusive and mutually beneficial in its future development. Today, China is a strong advocate of open sources, the development of open sources and the promotion of open sources. He mentioned that some open-source communities led by Chinese companies had flourished internationally, during the AI era..

In October 17th, this September, the China 718B big model became the focus of the industry, with a training philosophy based on data-free, focused thinking, in the SuperCLUE list to the third open source model. China officially announced yesterday that open Pangu-Ultra-MoE-718B-V1.1 is officially open on the GitCode platform, with full disclosure of model weights and technical details. Hardware requirements: Atlas 800T A2 (64GB, > = 32 carats), support naked machines or Dock..

August 12, 2011 - Kunlun World Wide's SkyWork AI Technology Release Week was launched on August 11, with one model released every day for five consecutive days, covering the core multi-modal AI scenarios. Today, Kunlun brings the upgraded version of Matrix-Game interactive world model in Matrix series of self-developed world model - "Matrix-Game 2.0", which is claimed to be a world model that realizes interactive real-time long sequence generation in general scenarios. In order to promote the development of interactive world modeling, "Matrix-Game 2.0" is said to realize the interactive real-time long sequence generation world model in general scenarios.

August 6, 2011 - OpenAI announced the release of two open source models, GPT-oss-120b and GPT-oss-20b, the first open source language models to be released by OpenAI since the 2020 release of GPT-2, both of which can be downloaded for free from the Hugging Face online developer platform. Both models, which are available for free download on the Hugging Face online developer platform, have been described by OpenAI as performing "at the cutting edge" in a number of benchmarks used to compare open source models. According to OpenAI, the Gpt-oss-120b model has been shown to perform "at the cutting edge" in core reasoning...

August 2, 2011 - Technology media outlet WinBuzz published a blog post yesterday (August 1) reporting that a mysterious AI model called Horizon Alpha has emerged on the OpenRouter platform, quickly topping the EQ-Bench Creative Writing charts, and is predicted to be the first open-source AI model for OpenAI. Horizon Alpha was released on July 31st, and has achieved excellent results on the EQ-Bench charts, surpassing Kimi K2 ...

July 29 news, yesterday, Smart Spectrum released a new generation of flagship model GLM-4.5, specifically for intelligent body applications to build the basic model. Hugging Face and ModelScope platform synchronization open source, model weights follow the MIT License. 1AI with the official introduction of the main points are as follows: GLM-4.5 in the inclusion of reasoning, code, intelligent body of the comprehensive ability to achieve Open source SOTA, in the real code intelligent body of artificial comparison assessment, the actual test of the best in China; the use of hybrid expert (MoE) architecture, including GLM-4.5 ...

July 19, 2011 - LMArena, the leading ranking of big models, has released its latest rankings, with the recently released Kimi K2 taking the top spot over DeepSeek R1 for open source models. According to LMArena, Kimi K2 took fifth place in LMArena's overall ranking based on its performance and 3,000 community votes. It's worth noting that Kimi K2 and DeepSeek R1 are the two Chinese models in the top 10 of the LMArena charts, but...

Finally waiting for the Kontext open source model. The model's API has been out for some time, the effect is very good. Because the effect is very good, so the feeling of open source hopeless, did not expect an unexpected surprise, open source model suddenly came. This should be the most playable image editing model in recent times. First, a quick look at fluxKontext What are the features: From this highlight, the function of this model is very clear, it is rushed to the intelligent P picture. Here's a quick preview of the effects and application scenarios: Style Transform: Cue: Transform to 196...

June 11 (Bloomberg) -- There's been a new tweak to OpenAI's open source model release schedule. In a post on the X platform on Tuesday, the company's CEO Sam Altman announced that the open-source models, which were originally planned for early summer of this year, are expected to be delayed until late summer and won't be available to the public in June. 1AI notes that in the post, Oltman wrote, "We're going to take a little more time for the open source weighting model, which means that people can expect it in late summer this year, but it won't be in June." He also mentioned that OpenA...

April 30 - Freepik, in partnership with AI startup Fal.ai, has launched its latest open-source AI image model, F Lite, which is unique in that the model is based entirely on the company's in-house dataset of about 80 million commercially licensed and secure content (SFW). Freepik's move circumvents common legal risks and ensures the copyright security of the model's output. Officially, the model utilizes 10B parameter diffusion to provide users with a high-quality image generation experience. F Lite is available in two versions: the standard version (...

April 14, at the recently held TED 2025, OpenAI CEO Sam Altman said, "The emergence of DeepSeek has not affected the growth of the GPT, and better open source models will be introduced." Market research organization App Figures recently reported that in terms of global non-game app downloads in March 2025, ChatGPT overwhelmed Instagram and TikTok with 46 million times, becoming the world's most downloaded (only counting Apple App Store...

Tencent HunyuanDiT recently released three new controllable plug-ins ControlNet with the community, namely tile (high-definition magnification), inpainting (image restoration and expansion) and lineart (line draft raw image), further expanding its ControlNet matrix. The addition of these plug-ins enables the HunyuanDiT model to cover a wider range of application scenarios, including art, creativity, architecture, photography, beauty and e-commerce, etc. 80% cases and scenarios provide more accurate...

Meta recently released a new generation of open source model series Llama3.1, including a 405B parameter version, whose performance is close to or even surpasses closed source models such as GPT-4 in some benchmarks. Llama3.1-8B-Instruct is an 8B parameter version in the series, supporting English, German, French, Italian, Portuguese, Spanish, Hindi and Thai, with context length up to 131072 tokens, and the knowledge deadline is updated to December 2023. To enhance Llama3.1-8B-Instr…

Llama3.1 has been leaked! You heard it right, this open source model with 405 billion parameters has caused a stir on Reddit. This may be the open source model closest to GPT-4o so far, and even surpasses it in some aspects. Llama3.1 is a large language model developed by Meta (formerly Facebook). Although it has not been officially released, the leaked version has caused a sensation in the community. This model not only includes the base model, but also benchmark results of 8B, 70B, and 405B with the largest parameters. Performance comparison: Ll…