ByteDance

February 17 news, byte jump in January this year released a new AI programming product for developers Trae, which is an AI Chinese IDE tool, but at that time only Mac version. According to the feedback of 1AI partners, the tool is now online Windows version, supporting Windows 10 and Windows 11 systems. Trae is powered by SPRING PTE, a Singaporean company under ByteDance, and supports AI Q&A, code auto-completion, Agent-based ...

February 12, 2011 - The Beanbag Big Model team announced today that the Byte Jump Beanbag Big Model team has proposed a new sparse model architecture, UltraMem, which effectively solves the problem of high access memory in MoE inference, improves inference speed by 2-6 times compared to the MoE architecture, and reduces the cost of inference by up to 83%. The study also reveals the Scaling Law of the new architecture, proving that it not only has excellent Scaling characteristics, but also outperforms MoE in terms of performance. The study also reveals the Scaling Law of the new architecture, demonstrating that it not only has excellent scaling characteristics, but also outperforms MoE. Experimental results show that the training scale of 20 million value...

January 22 news, byte jump's Doubao big model 1.5 Pro officially released today, the new Doubao-1.5-pro model enhanced comprehensive ability, in the knowledge, code, reasoning, Chinese and other benchmarks, the comprehensive score is better than the GPT-4o, Claude 3.5 Sonnet and other models. At present, Doubao-1.5-pro has been grayed out on line in Doubao App, while developers can also directly call the API in the volcano engine. According to the official introduction of the byte jumping, Doubao big model 1.5 P...

ByteDance recently launched Trae, an AI integrated development environment (IDE) for Chinese developers, which aims to compete with internationally recognized tools such as Cursor and Windsurf, especially in terms of the experience of Chinese developers, and hopes to fill the shortcomings of existing tools. In the daily programming, many developers often need to switch between English and Chinese, and although this mixed working mode is commonplace, it is inconvenient for many Chinese developers. Existing IDEs such as Cursor, Windsurf and ...

January 17 news, byte jumping beanbag January 16 announced that beanbag computer version and web version of the new online AI programming function. This function supports uploading multiple local code files with one click, introducing GitHub open source repository in real time, and quickly obtaining the complete context of the project without copying the code segment by segment. 1AI learned from the official introduction that Doubao's new AI programming function is equipped with a brand new code editor that supports immersive reading; code snippets that need to be interpreted or adjusted can be accurately circled; and the directory can be easily switched when viewing the code repository. In addition, users can also by beanbag to...

Byte jumping's Doubao big model announced on December 30 to the public the technical progress of the whole field in 2024, claiming that since its debut on May 15, 7 months ago, the model ability in general language, video generation, speech dialogue, visual understanding and other aspects has crossed the first echelon of the international. According to reports, as of December 2024, the latest version of Doubao's general-purpose model Doubao-pro-1215 has improved its comprehensive capability by 32% compared to May, and has been fully aligned with GPT-4o, and the effect of some complex scenario tasks such as math and specialized knowledge...

On December 18th-19th, the Volcano Engine FORCE Power Conference was officially held in Shanghai. The most interesting point of this conference is undoubtedly the upgrading of the beanbag model family and the release of the new beanbag visual understanding model. The Beanbag Visual Understanding Model has industry-leading content recognition, understanding and reasoning capabilities, as well as more detailed visual description capabilities. It is equivalent to adding a pair of eyes to the big model so that it can recognize the real world and understand the real world like humans. This undoubtedly expands the forms and application scenarios of human interaction with big models. Just as the volcano engine ...

At the Volcano Engine Force conference, ByteDance officially released the Beanbag Visual Understanding Model, which provides enterprises with multimodal large model capabilities. Thousands of tokens input price is only 3 cents, a dollar can deal with 284 720P pictures, the official claim is 85% cheaper than the industry price. In this regard, Li Liang, vice president of Jitterbug Group, said in an article today: This is not a price war. Beanbag big model through technological innovation to reduce costs, in the algorithm, software engineering and hardware solutions to do a lot of optimization, 3% / thousand tokens pricing also has considerable gross profit. And...

On December 18, ByteDance's Dream AI announced that its new image model 2.1 was officially launched. It supports the rapid generation of Chinese and English image posters by controlling the color and position of text and other elements through simple commands. It is called "generating posters in one sentence". The official Dream AI provides a common Prompt template, i.e. Artistic Style + Body Description + Text Layout: Artistic Style: Determine the overall tone of the image, such as realistic style, illustration style, or more specific styles, such as Chinese ink and watercolor style, etc. Description of the main body of the image: Focus on the main image content, for example, describing the poster's ...

December 18 news, 1AI from the byte jump official learned that in today's Volcano Engine Force Conference, byte jump officially released beanbag visual understanding model, to provide enterprises with multi-modal large model capabilities. Beanbag visual understanding model thousand tokens input price is only 3%, one dollar can handle 284 720P pictures, the official claim than the industry price is 85% cheaper. Beanbag 3D generation model is also officially unveiled in this event. Combine it with the volcano engine digital twin platform veOmniverse, can complete the smart...

December 13 news, Peking University and byte jumping on December 12, the establishment of a joint laboratory of beanbag big model system software. The scientific research of the joint laboratory will be based on ByteDance's big model of beanbag. Researchers from universities and enterprises will focus on training, reasoning and other scientific research work around the big model in the laboratory. With the construction and development of the laboratory, there will also be a batch of Beihang University students entering the laboratory for internship and practice, and carrying out scientific research with front-line researchers in the industry. Note: Byte Jump Beanbag Big Model team was founded in 2023, and its research direction covers deep learning,...

According to "intelligent emergence" today cited sources familiar with the matter, byte jumping has elevated the product priority of the dream, trying to use a new path to create "AI era jittery voice". Sources familiar with the matter said that this move is because the management of ByteDance judged that the "AI dialogue class" products may only be the "intermediate state" of AI products, and that a more ideal product form in the long term will probably require a more visual user experience and a lower threshold for user use. Currently, Doubao's user activity is 2-3 days per week, and the number of times a user sends a message is only 5-6 times per day, with a single time of about 2 minutes, and the per capita user time is only ...

Byte jumping video generation model PixelDance has officially opened internal testing in the computer version of the beanbag, some users have opened the experience portal. The internal test page shows that users can generate ten videos per day for free. According to 1AI's previous report, the PixelDance video generation model was first released at the end of September, and the earliest through the dream AI, volcano engine for creators and enterprise customers to invite a small-scale test. According to early internal test creators, when PixelDance generates a 10-second video, the best results are achieved by switching the camera 3-5 times, and the scenes and characters can maintain a good...

ByteDance founder Zhang Yiming has been actively involved in the group's AI strategy even though he has stepped down as CEO, insiders at ByteDance said, according to a Dec. 8 report in the Financial Times. Zhang personally oversaw the recruitment of AI engineers and researchers from rival companies, the people said. Zhang has spoken internally about his goal of "universal AI," a system with human-like intelligence. The report also mentioned that some industry insiders are skeptical of this goal, as the company has made several failed bets on the future of the technology. He has previously traveled ...

The former Ali Tongyi Qianqian big model technology leader Zhou Chang (fancy name: Zhong Huang) was exposed on July 18th will leave to start a business, but on October 23rd, there is news that Zhou Chang has low-profile to join the byte jump. On November 13, a news about Alibaba's decision to apply for arbitration "Tongyi big model former employee Zhou Chang violated the non-compete agreement" quickly spread in the AI circle, close to Tongyi's industry insiders said "the situation is true". According to the first financial report today, a service byte headhunting company staff revealed: "Resignation to start a business is a front, is afraid of being non-compete, but...

ByteDance's intelligent AI assistant Doubao has now upgraded its text-to-picture capability to support one-click generation of specified text. Users can add text requirements to the prompt words to generate pictures with specified text. At present, this function has been tested in the Beanbag App, that is, Dream has also been tested on a small scale. According to "Tiger Sniff" reported today, Doubao big model team responsible person said, Doubao text generation model through the LLM and DIT structure, with better native Chinese data learning ability, and on this basis, strengthened the ability to generate Chinese characters, greatly improving the generation effect. Doubao's webpage and computer version...

December 4, according to the news of interface news on the 3rd, byte jumping's AI assistant Doubao is now online picture understanding function, Doubao App and Doubao PC new photo and camera buttons, after uploading pictures can recognize the content of the picture. Doubao's "Picture Understanding" function can recognize the elements and basic features contained in a picture. Users can also ask where an attraction is located or which movie or TV show a character is from. In addition, not long ago in November, Doubao launched a video generation internal test function, support for picture and text a key to become a movie, dynamic operation of the mirror and multi-camera a...

Byte Jump's Dream AI has announced that two major video models have been officially launched in full volume. Officially, it contains S2.0 Pro with first frame consistency and P2.0 Pro video generation model with very high cue word compliance. At present, you can open the Dream App or visit the Dream website to select the video S2.0 Pro or video P2.0 Pro generation model. S2.0 Pro official said, S2.0 Pro in the i2v (IT home note: image to video, graphic video) scenarios have a very good first frame ...

Byte jumping beanbag big model team yesterday announced universal image editing model SeedEdit, users only need to input simple natural language, you can carry out diversified editing operations on the image, including retouching, dressing, beautification, style transformation, and adding or deleting elements in the designated area, and so on. Officially, the model has already been tested on the PC side of Doubao and the web side of Imagine Dream. The official said that SeedEdit can widely adapt to the diversified editing needs of different users, with strong controllability, good editing effect and no sense of sticker, which is mainly reflected in three aspects: High-precision command comprehension: in both Chinese and English, and proper nouns...

Recently, namely dream AI announced that from now on byte jumping self-developed video generation model Seaweed for platform users officially open. After logging in, under the "Video Generation" function, users can experience the video model by selecting "Video S2.0". Seaweed video generation model is part of the beanbag model family, with professional-grade light and shadow layout and color harmonization, the visual sense of the screen is very beautiful and realistic. Based on the DiT architecture, the Seaweed video generation model can also realize smooth and natural large-scale motion pictures. Tests show that the model only needs 60s to generate a high-quality AI with a duration of 5s...

Byte jumping's AI content platform, namely dream AI, announced on November 8, byte jumping's self-developed video generation model Seaweed is officially open to platform users. After logging in, under the function of "video generation", users can experience it by selecting "Video S2.0" for the video model. The attached link is as follows: https://jimeng.jianying.com/ai-tool/video/generate At the end of September, ByteDance officially announced its entry into the field of AI video, releasing two video generation models of the beanbag model family...

November 4 news, byte jumping quietly launched an AI model sharing community called "stove rice Lumi", the platform is said to integrate model sharing, Workflow building and LoRA training and other functions, users can upload their own development of AI models, and share them with others in the community, the platform also supports users to build AI workflow. The platform also supports users to build AI workflows. However, it is noted that the platform is still in the internal testing stage, and at present the mainstream search engines have not included the relevant website, but you can use the "artistrylab.net" web address to directly...

HybridFlow is the result of a joint research project between ByteDance's Beanbag Big Model team and the University of Hong Kong. Officially, HybridFlow (open source project name: veRL) is a flexible and efficient RL training framework for big models, compatible with a variety of training and inference frameworks, and supporting flexible model deployment and Multiple RL algorithm implementation. The framework uses a hybrid programming model that combines the flexibility of Single-Controller and the efficiency of Multi-Controller to better realize and execute multiple...

Byte Jump released a factual clarification announcement about interns destroying big model training". Byte Jumping big model training was attacked by interns, and the person involved has been dismissed The announcement said that recently the media said that Byte Jumping big model training was attacked by interns", and after the company's internal verification, there was indeed a serious disciplinary violation of the interns of the commercialization technology team, and the interns have been dismissed. Byte Jumping said that the relevant reports also exist in part exaggerated and inaccurate information, the special explanation is as follows: First, the intern involved maliciously interfered with the model training task of the commercialization technology team's research project, but it did not affect the commercialization of the official project and the online business, and also...