free and open source

On October 23rd, MiniMax rare technology announced today the model weight of MiniMax M3.1 Flash, total parameter 428B, active parameter 23B. The M3.1 Flash model capabilities were fully enhanced compared to the previous M3 model. In addition, M3.1 Flash further enhances the reasoning efficiency through the optimization of the whole model by thin attention, quantitative sensory training and Power-aware learning: whole model thin attention: first three full notes..

On October 20th, the Ant Ultimate announced today the Open Source Ming-3.0 Voice Series, which provides voice synthesis, audio understanding and dialogue, and voice recognition series of five models. Developers can choose between tasks or combine different competencies into their own work streams. The Voice Design of the Ming-3.0-ttts series can describe the target sound in text while also supporting the audio setting of the reference audio. More specifically, the creator can use the director's mode to describe the way the lines are performed, how emotions change and what words need to be read again..

On October 17th, the information mixer was released on October 16th with Hy Memory 2.0, along with a dedicated open source memory model, memory frame and MemHub application. Three layers of capability can be used independently and can be combined: the model is responsible for memory reading and writing, the framework manages the storage, updating and recall of memory, and MemHub will access different AI tools. The framework adds working and code memory to the chat memory, recording mission objectives, identified practices, stage results, and reusable experiences and applicable conditions. Code scene saves version evolution..

On October 16, local time, on October 15, Meta released the Open weight AI model Muse Spark Open. Meta officially claims that Muse Spark Open is the most capable American open weight model to date. Muse Spark Open supports the context window of 1M token by using a primary multimodular structure with a total parameter volume of 816B (with activated parameter 42B). The model is designed for code construction in collaboration with intelligent bodies to handle long-range tasks..

On October 1st, the Index LLM team officially launched the Index-Translate model on September 30, with 2B / 9B / 35B-A3B (preview) text model weights in Hugging Face and ModelScope. Index-Translate is a multilingual translation model family based on Qwen3.5. Text model covers 150 languages, including English, supporting terminology, format, security..

On September 22nd, Mi was officially released yesterday and opened with the Xiaomi MiMo-V2.6 series, which contains both Pro and Flash original whole-state models. MiMo-V2.6-Pro received 46 points in Artificial Analysis Inteligence Index, more than Kimi K3 and Qwen3.8 Max, but closed-source models such as Claude Table 5.1, GPT-6 Astra still..

On 21 September, according to the news, ZCode announced today that the official restructuring had been completed in response to community feedback on the safety of ZCode products and apologized to all users. The ZCode Open Source (https://github.com/zai-org/ZCode) has been put under community supervision to make ZCode open and transparent. Thank you very much to the community developers who found out about the ZCode problem, and then we'll set up a normal product safety gap mechanism, and we'll welcome ongoing checks and feedback from the developers' partners..

On September 21st, Ali thousands announced yesterday that the open source image series is currently a Qwen-Image-2.1 model of open source images that combines the effects of generation, the efficiency of reasoning and the cost of use. Qwen-Image-2.1 integrates the texture map with image editing in the same model, with only 7B parameters in the visual generation component, and the original supports the generation and editing of transparent images. Officially, the main features of this update include four aspects: a small model that is strong and highly priced: a light model structure that optimizes reasoning, balancing quality generation and costing. It's raw..

On September 20, Chinese telecommunications released on September 17, a new-generation star-size model, Xing4.0-29B-A4B, with a total of 29B, a activation parameter of only 4B, with original support of 256K in the context, which can be extended to 512K, the first large-scale model in the country to optimize the depth of complex engineering tasks based on national capacity and framework. Officially, from training chips to reasoning deployment, the whole of the country's production — training, adaptation of the national production framework, self-study structures, open-source ecology — is one route. From Super..

Hypit is an open-source AI video creation tool, with a "repeated explosive." Users need only provide a reference video that Agent can automatically decipher its structure, graphic rhythm and subtitle styles and generate editable project files. When the text is modified, the sound, subtitles and animations are automatically synchronized, replacing the anchor, product and language, and producing hundreds of variants in bulk. Hypit supports access to multiple models, which are often used for advertising, electrical bringer and content matrix production, in conjunction with Claude Code, Codex, Bean Pack, etc., Agent. Hypit function..

On September 19th, Mini Max has opened the source code for Mini Max Code CLI and the current warehouse corresponds to 0.4.12 source preview, the first code defaults on the MIT license. The project covers terminal interfaces, headless execution mode and Agent Clinic Protocol, readable and modified codes, running commands and tests, and also supports search, plugins and multimodular tools. Developer can log in with Mini Max account, or via BYOK access and..

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 September 6th, Ali thousands asked for Qwen-Drive-1.0-4B, an automatic driving visual language model based on Qwen3.5-4B. Officially, Qwen-Drive-1.0 is the first visual language base model for auto-driving, which harmonizes 3D perception and visual questions and answers at the pre-training stage and further extends to sports planning, while retaining the original VLM structure for pre-training completely. The model uses a phased training programme that combines driving supervision with universal visual language data to obtain driver..

Deep Face Cam is a local face exchange desktop application that is compatible with MacOS and Windows. Images, videos and real-time cameras are supported for face-to-face change, and all processing is done locally at the equipment to ensure data privacy. Provide transparent and manageable use experience in an open-source, auditable manner. Deep Face Cam Local Data Processing: All face replacements are performed locally, media files need not be uploaded to ensure data privacy. Open source audit: Full source code is publicly available, following AGPL-3.0 protocol, free to check, build and modify. Model..

On September 1st, the S.I.S.S.S.S.-Syllable Fire announced today the launch and opening source of X2.5-4B and X2.5-1.7B generic large models at the end. It was stated that the two models had original support of up to 1 million Token context windows and were the only products in the current end-side model that supported that level of context. The two models use a mix of attention structure and optimize around core competencies such as intelligence, code, mathematics and command compliance. Based on the National Calculator Platform, completes the whole process, using about 20 trillion Token to diversify into..

On August 29th, a new generation of flagship models was released and opened. The model uses a hybrid specialist structure with a total parameter of 770B, a activated parameter of 49B, a context length of 1M, and focuses on productivity enhancement tasks such as software engineering, game development, office and scientific research. 163 internal experts completed 203 project blinding, Hy4 preview average score 2.99 / 4.00; the government also demonstrated that the model was produced through MCP operation Unreal 5..

On August 27, the first original multi-modular model of the GLM-5 series, GLM-5.3-Flash (320B-A18B) came on line last night and was open. The total size of the model is 320 B (320 billion) and only 18 B (18 billion) is activated. Before it was officially released, a large-scale test was carried out on OpenCode and OpenRouter with an anonymous model, Ox-Alpha. Ox-Alpha quickly became the most popular model of the week, creating..

On August 21st, Comtech announced today an open-source lightweight, original unified multimodular model Sensenova U1.5 Lite. It was described that Sensenova U1.5 Lite had a background length of 3-4k, which simultaneously addressed multiple constraints such as subject, quantity, spatial relations, text, layout, style, etc., and enhanced the stability of complex visual tasks. In addition, Sensenova U1.5 Lite has the following characteristics: higher quality visual production: improvement of the overall image, colour, material, light..

SpeakoFlow is a free open-source, cross-platform transliteration tool. It allows the user to perform a hearing, translation, colouring and AI assistant with voice in any application at a speed of more than 150 words per minute. Default local processing, open source customization is the ideal option for efficiency and privacy. The SpeakoFlow function Applies any hearing and writing: Press the button and speak, text is entered in real time into any mail, editor or chat box that is used; supports more than 150 words per minute, 3 times faster than typing, with no hand-to-hand operation. Local privacy protection: voice..

On 29 July, according to TechCrunch, Anthropic CEO Dario Amodei published a book on local time on 27 July, responding to outside rumours that Anthropic supported a ban on open-source weight models. In Bo, Amodei made it clear that "Anthropic has never advocated the prohibition of open-source weight models" and stressed that open-source weight models that do not have a dangerous capability are "a public good" and can bring value to businesses, developers and researchers. He also pointed out that even if he came..

On July 28th, the topic of whether AI should open up weight has recently become the focus of industry discussions. Last Friday, more than 20 companies, including Ying Weida, Meta and Microsoft, signed an open letter entitled Open Weights and American AI Leadership in support of Open Weights. By the end of the week, OpenAI, Google and SpaceX also signed the letter. Anthropi..

News from July 26, Andrew Ng (24 July) released a dynamic in LinkedIn, announcing the launch of OpenWorker, an open-source AI smart body, not only for chatting, but also for the efficient performance of the client's task. On the running platform, OpenWorker for Apple MacOS is on line, and the Github page shows Windows versions supported but not yet signed. OpenWorker based on the original..

On July 23rd, Inweida CEO Wong In-hoon, in an interview with Axios, stated that the United States had no fear of China’s open source AI models, and should be more vigilant in demanding that they be banned. He called the Chinese models "excellent" and that good open source models should be used, as should United States enterprises. This statement follows the release of Kimi K3 on the dark side of the month, and the renewed discussion of China’s open model risks in American politics. OpenAI and Anthropic accused Chinese competitors of distilling their model capabilities and alerting Washington to..

On July 8th, the Hong Kong Big Data Intelligence Laboratory (HKUDS) recently released the Open Source AI Smart Body Mentor DeepTutur v. 1.5, which completed several upgrades on document decomposition engines, RAG intake pipe lines and Python 3.14+ compatibility. DeepTutor's core design is to integrate the six components of chat, testing, research, visualization, problem resolution and learning path planning into the same smart body closed loop, so students do not need to interrupt backstage engines when they switch between assignments, and learning context and multilevel memories automatically coincide..