Model performance

With the rapid iteration of generated AI products such as ChatGPT and Copliot, the demand for training data is growing exponentially, and it is also one of the key links to improve the performance of the large model. In order to solve the problem of training data shortage and poor quality, Microsoft Research Institute has released an AIAgent - AgentInstrument, which is specially used to generate high-quality synthetic data. Orca-3 also surpasses other models such as LLAMA-8B-instrument and GPT-3.5-turbo in many indicators

Special statement: The label name and page content of this page are automatically generated by the website system for the classification of information content, and are only used for indexing information content, so as to facilitate users to index relevant information reports. If the label name involves trademark information, please visit the trademark brand official for details. Please do not use the content of the label page of this website as reference information. This website does not have any association with the possible trademark name information. We will not be responsible for any errors, inaccuracies or omissions caused by the content of this page. The webmaster's home will try its best to ensure the accuracy and reliability of the information provided, but it does not guarantee the accuracy and reliability of the relevant information. Readers should further verify before using it, and be responsible for any self determined behavior. Any unit or individual who thinks that the content of this page may be suspected of infringing their intellectual property rights or have false content can timely submit a written notice of rights or false statement to the webmaster's home, and provide proof of ownership and detailed proof of infringement or false information（ Click to view the feedback contact address ）。 After receiving the above feedback documents, this website will verify the information according to laws and regulations, and immediately communicate to delete relevant content or disconnect relevant links.

And“ Model performance ”Related hot search words of:

Related“ Model performance ”11970 pieces of information

[Tencent Cloud] 11.11 Cloud Shenghui! The first year of ECS is 1.8% off, and you can get 3 months free if you buy one year!

11.11 Cloud Shenghui! Massive products, easy to cloud! The first year of ECS is 1.8% off, and you can get 3 months free if you buy one year! Super value offers and stable performance make your cloud journey more enjoyable. Come to Tencent Cloud to buy!

Docker container image
Go to see

Docker container image 60 yuan/15 days

Combination purchase of popular products
Go to see

Combination purchase of popular products As low as 1 yuan

Tencent Cloud x NVIDIA Acceleration Plan
Go to see

Tencent Cloud x NVIDIA Acceleration Plan Up to 100000 yuan of support fund

2-core 2G ECS
Go to see

2-core 2G ECS 112 yuan/year

See more information>>

Tencent Cloud 12-20

advertisement
Recommendation Microsoft uses AI Agent to generate 25 million pairs of high-quality data, greatly improving the performance of the large model!

With the rapid iteration of generated AI products such as ChatGPT and Copliot, the demand for training data is growing exponentially, and it is also one of the key links to improve the performance of the large model. In order to solve the problem of training data shortage and poor quality, Microsoft Research Institute has released an AIAgent - AgentInstrument, which is specially used to generate high-quality synthetic data. Orca-3 also surpasses other models such as LLAMA-8B-instrument and GPT-3.5-turbo in many indicators.

Microsoft
Intel was among the first to pass the performance test of AISBench large model! Generation 5 Xeon can reach 2493 tokens per second

Recently, the fifth generation Intel Xeon processor passed the performance test of AI server system organized by the China Institute of Electronic Technology Standardization with excellent performance. Intel has also become one of the first enterprises to pass the reasoning performance test of AISBench big language model. At the same time, it also highlights that Intel Xeon can provide enterprises with the function of "out of the box", that is, it can deploy a part of AI workload on common systems, thus bringing better total cost of ownership advantages to customers.

Intel Xeon processors AISBench test AI server
Apple AI amplification move! Claim that its device side model performance is better than GPT-4

In a recent paper, Apple's research team announced that they have proposed a model ReALM that can run on the device side, and this model can surpass GPT-4 in some aspects. The parameters of ReALM are 80M, 250M, 1B and 3B, which are very small and suitable for mobile phones, tablets and other devices. This research is expected to be used to improve the Siri assistant on Apple devices and help Siri better understand and process the context in user queries.

Apple ReALM AI
IDEA proposed to improve the performance of the ToG mind map big model by 214%!

The Think on Graph technology, launched by IDEA Research Institute, Microsoft Asia Research Institute, Hong Kong University of Science and Technology and other research teams, has set off an innovation in the field of deep reasoning. This technology successfully makes up for the weakness of the big model in terms of illusion problems in the fields of finance, law, medicine and so on by tightly coupling the big model with the knowledge map. This research result brings new ideas and methods to the development of deep reasoning field, and provides strong support for realizing more reliable and interpretable reasoning ability of large models in various fields.

ToG
ROUGE matrix: large model performance evaluation tool

The ROUGE matrix is a tool for evaluating the performance and capabilities of large language models. In the field of natural language processing, evaluating the performance of models is very important. However, due to the unpredictability of language output and the complexity of human language, traditional machine learning evaluation methods have limitations. ROUGE matrix is an important tool to evaluate the ability of language model summarization. Through continuous improvement and combination with other indicators, the performance of language model can be more accurately evaluated.
LLM evaluation test framework DeepEval can evaluate the performance of large models offline

DeepEval is a framework for evaluating and unit testing language model applications. It provides various indicators to test the performance of responses generated by language model application in terms of relevance, consistency, unbiased and non-toxic. If you are developing chat robots and language model applications, DeepEval is definitely a good helper to improve engineering efficiency.

DeepEval
Recommendation The performance of 20B large model is comparable to Llama2-70B! Fully open source, from the base to the tools

[New Zhiyuan Guide] The record of domestic model open source has been broken again! InternLM-20B, which is open sourced by Shanghai AI Lab and other institutions, can even draw with Llama2-70B? Just now, the domestic open source model parameter record has been refreshed! On September 20, Shanghai Artificial Intelligence Laboratory and Shangtang Technology, together with the Chinese University of Hong Kong and Fudan University, officially opened the InternLM-20B model with 20 billion parameters. Project address: https://github.com/InternLM/InternLM Mota Community: https://modelscope.cn/organization/Shanghai_AI_Laboratory The 20 billion parameter version of the Scholar Puyu model can be said to be "no price added", with less than one third of the parameters, but its performance can challenge the benchmark of today's open source model - Llama2-70B. The "timely help" of Shanghai AI Laboratory will certainly bring great value to the large model in the field of landing.

Model
Zuckerberg's Magnification: New AI big model performance upgrade several times to be the world's strongest | AI outpost

Phoenix Science and Technology News AI Sentinel On September 11, Beijing time, Facebook founder Mark Zuckerberg has targeted OpenAI. People familiar with the matter said that Meta, the parent company of Facebook, is developing a new AI system, hoping that it can be as powerful as the most advanced model provided by OpenAI. AI Sentinel of Phoenix Network Technology will continue to pay attention to this.
SPP open source: support multi person collaboration to improve the performance of large models

SoloPerformancePrompting is a multi person collaboration assistant designed to improve the performance of large language models. It solves tasks in a collaborative way through multi person settings. SoloPerformancePrompting can significantly improve the performance of large language models on various tasks by introducing the method of multi person cooperation, providing users with better experience and assistance.
Tsinghua releases the performance report of the large model: Baidu is the first person who understands Chinese better in GPT-4

The AI big model has become the hot spot this year. OpenAI's ChatGPT has taken the lead, followed by Microsoft, Google, and Meta. The domestic technology giants have also quickly followed. Now it is said that there are hundreds of big models in China. What is the strength of so many models? A few days ago, the School of Journalism and Communication of Tsinghua University released the Comprehensive Performance Evaluation Report of the Large Language Model, comparing the large models at home and abroad. This is related to the fact that Baidu's big model contains a large number of Chinese texts, so it can better handle the content related to local culture.
Huawei Cloud Research Team proposes PanGu Coder2 model and RRTF framework to improve the performance of large models

Researchers from Huawei Cloud, Chinese Academy of Sciences and Peking University recently proposed a new framework called RRTF, which can effectively improve the code generation performance of large pre trained language models. The RRTF framework improves the performance of code generation LLM through natural language LLM alignment technology and ranking feedback. 6. In addition to improving code generation efficiency, the research team also proposed the optimization method used by PanGu Coder2 to ensure rapid reasoning, which is very important for practical applications.

Large model AI Headlines
Tencent launched a new generation of large model, "Hybrid Turbo", with significantly improved performance and 50% lower price

On September 5, Tencent announced the launch of a new generation of large model "Hybrid Turbo". Compared with the previous model, Tencent's hybrid Turbo has significantly improved its performance, with training efficiency increased by 108%, reasoning efficiency increased by 100%, reasoning cost reduced by 50%, and decoding speed increased by 20%. The effect of benchmarking GPT-4o on multiple benchmark tests, and third-party evaluation ranked first in China. At the Tencent Global Digital Ecology Conference, Qiu Yuepeng, Vice President of Tencent, COO of Cloud and Smart Industry Business Group and President of Tencent Cloud, announced that the Tencent Hybrid Turbo was launched in the Tencent cloud, and the input and output prices were only half of the previous generation models. Enterprises and developers can access and use it directly on the cloud. Tencent's collaborative SaaS products are fully integrated into Tencent's hybrid model.

Tencent Hybrids Turbo Large language model Cloud computing service
Recommendation Performance kills SD3, DALL · E-3, and open-source cultural map model kills big dark horse

Midjournal just made a big update yesterday. Today, a big dark horse - FLUX. 1 - has emerged in the open source field of Wensheng pictures. According to its test data, its performance greatly exceeds that of the DALL · E-3 and MidjourneyV6 closed source models, and the Ultra, Medium, Turbo and SDXL of the open source SD3 series are all killed. The image quality, details and lighting effects generated by FLUX. 1 are not bad. Let's look forward to his Wensheng video model.

FLUX.1
Recommendation AI Daily: Performance exceeds GPT-4o! The Meta405 billion parameter open source model Llama3.1 was leaked; Stability AI open source audio generation model Stable Audio Open

Welcome to the column of [AI Daily]! Here is a guide for you to explore the AI world every day. Every day, we present you hot content in the AI field, focus on developers, and help you to understand technology trends and innovative AI product applications. Click to learn about new AI products: https://top.aibase.com/1 , shocked the AI world! 14. Microsoft Research Institute promotes AI framework E5-V: Simplify multimodal learning with text pair single-mode training to reduce costs The E5-V framework jointly launched by Microsoft Research Institute and Beijing University of Aeronautics and Astronautics simplifies multimodal learning through single-mode training to reduce costs, shows excellent performance in multiple tasks, and represents significant progress in multimodal learning.

Llama3.1
Alibaba Cloud comprehensively upgrades GPU cloud services: reasoning performance of AI big model can be improved by 100%

Alibaba Cloud launched a fully upgraded GPU cloud service. Through the new optimized GPU suite, the performance of multi GPU reasoning services can be improved by up to 100%. For the container technology most commonly used for calling GPU computing power, Alibaba Cloud launched the ACK Cloud native AI suite to help developers and enterprises develop and deploy AI applications more efficiently and accelerate the implementation of large model applications. Xiaopeng Auto, Lenovo, Deloitte, Weibo, Perfect World, Himalaya and others have been connected to Alibaba Cloud's Tongyi model to provide consumers and customers with rich AI application services on Alibaba Cloud.
Only 3.8B parameters are activated, and the performance is comparable to that of the same 7B model! Training fine-tuning can be used, from Microsoft

Only 60% of the parameters need to be activated to achieve the performance equivalent to the fully activated dense model. A new research by Microsoft Research Asia has realized the full sparse activation of the model, which has greatly reduced the reasoning cost. At a large scale, the sparse activation model is likely to achieve the same performance as the dense model, which provides a useful reference for designing and training large-scale sparse activation models.

Microsoft
Recommendation Nvidia has 340 billion yuan of open source, and 98% of synthetic data has trained the strongest open source universal model! Performance benchmarking GPT-4o

[Introduction to New Zhiyuan] Just now, Nvidia's newly released open source model Nemotron-4340B may completely change the way LLM is trained! Perhaps all industries no longer need expensive real world data sets. It is worth noting that this advantage is mainly due to the low long/lengthy response rate.

Yingweida GPT-4o artificial intelligence
New performance breakthrough! Comparable to international standards! Domestic thousand calorie intelligent computing cluster helps domestic large model training and promotion

The new generation of AI technology represented by the big model is a strategic technology and an important driving force leading a new round of technological revolution and industrial change. With the increasing demand for computing power, AI computing base companies are increasingly becoming an important part of the new wave of AI technology. In May 2024, Moore Thread and SmartSpectrum AI carried out a new round of GPU large model adaptation and performance testing, including large model reasoning and large model pre training based on the Quae thousand Ka intelligent computing cluster. About Moore Threads Moore Threads is an integrated circuit high-tech company focusing on the design of full function GPU chips. It can provide a strong computing acceleration capability for a wide range of scientific and technological ecological partners, and is committed to building a metacomputing platform that provides multiple computing power for the next generation of Internet.

artificial intelligence technology Computing power demand GPU large model
AliCloud Tongyi Qianwen open source Qwen2 series models: significantly improved performance, including 5 sizes

Alibaba quietly released the Qwen2 series model, which is a major update after the Qwen1.5 upgrade. Qwen2 series includes five pre training and command fine-tuning models of different scales, namely Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B and Qwen2-72B. The Qwen team will also continue to train larger models, explore the ScalingLaw of models and data, and expand Qwen 2 into a multimodal model.

Alibaba Qwen2 series models Pre training model
Kunlun Wanwei Announces Open Source of 200 Billion Sparse Large Model Skywork MoE with Strong Performance and Lower Cost

Under the background of rapid development of large model technology, Kunlun Wanwei opened a landmark sparse large language model Skywork MoE. This model not only performs well in performance, but also significantly reduces the reasoning cost, providing an effective solution to the challenges posed by large-scale intensive LLM. The open source of Skywork MoE brings a powerful new tool to the large model community, which helps to promote the development of artificial intelligence, especially in scenarios where large-scale data needs to be processed and computing resources are limited.

Tiangong MoE Skywork-MoE AI Headlines
32 Expert MoE big model is free for commercial use! The performance of Llama3 is fully benchmarked, and the reasoning consumption of a single token is only 5.28%

Each token requires only 5.28% of the calculation force, and the accuracy can be fully calibrated to Llama3. Open source big model, add another heavyweight player - MoE model of 32 experts from Inspur Information, source 2.0-M32. From innovative research and development to open source, Inspur Information will stick to the research and development of basic big models, reduce the threshold for enterprise users to use big models, and accelerate the upgrading of industrial intelligence.

Large model Llama3
LLama3-V model officially released: performance is equivalent to GPT4-V and Gemini Ultra

The LLAma3-V model was officially released, which is a new SOTA open source VLM model. The LLAma3-V is not only better than the LLaVA in terms of performance compared with GPT4-V, GeminiUltra and ClaudeOpus, but also the model volume is only 1/100 of theirs. This news is very exciting. I believe that with the advent of LLAma3-V, it will bring new development and breakthrough to the field of artificial intelligence.

LLAma3-V model SOTA open source VLM Artificial intelligence
UIUC releases StarCoder2-15B-Instruct code model without OpenAI data on the performance list

Researchers from UIUC and BigCode recently released the StarCoder2-15B-Instrument code model, which has made a significant breakthrough in code generation. This model has successfully surpassed CodeLlama-70B-Instruct and topped the list of code generation performance. The open source dataset and training process of the model lay a solid foundation for future research and development in this field.

OpenAI Code big model AI Headlines
DeepSeek releases the price of cabbage with the performance of V2 model GPT-4

DeepSeek has made important progress in the field of open source MoE models. After opening the first MoE model in China in January this year, DeepSeek has released the second generation MoE model: DeepSeek-V2. This new model has shown excellent performance in many performance evaluations, competing with closed source models such as GPT-4, and has significant advantages in cost efficiency. DeepSeek provides an official dialogue website and an open platform. Users can open a free dialogue at chat.deepseek.com DeepSeek API open platform provides a large number of tokens for registration, so as to support the development of high-quality projects.

DeepSeek AI Headlines
Recommendation Apple launched the understanding and transformation model ReALM, with performance exceeding GPT-4

Apple researchers have introduced an innovative model ReALM, which can transform the reference parsing problem into language modeling problem, and can greatly improve AI assistants to handle many complex or ambiguous descriptions. We selected oranges, pears, apples, orange sodas, shampoos, short sleeves, slippers and other items when shopping online, and then told AI assistants to only settle for fruits. This module can also help the model to improve its analytical ability by self-learning through the user's feedback mechanism.

GPT-4 ReALM
Alibaba has opened a 100 billion parameter model Qwen1.5-110B, which outperforms Meta-Llama3-70B

Alibaba announced that it will open source its latest Qwen 1.5 language model - Qwen 1.5-110B. This is the largest model in the Qwen 1.5 series and the first model in the series with more than 100 billion parameters. It performs well with the recently released SOTA model Llama-3-70B, and is significantly better than the 72B model.
Synthesia will launch a new digital avatar model EXPRESS-1 with performance exceeding HeyGen

Synthesia， A leading AI company recently announced that it will soon launch a new digital avatar model, EXPRESS-1. The launch of this new model will undoubtedly further promote the application of AI in the field of video production. We are looking forward to the launch of this new model and the new video viewing experience it will bring us.

Synthesia AI Headlines
Recommendation Open source big model Llama 3 king returns! The maximum bottom card is 400 billion, and the performance is close to GPT-4

Zuckerberg: "With Llama3, the world can have the most intelligent AI." ChatGPT kicked off the big model competition, and Meta seems to be coming from behind. Maybe only OpenAI's GPT-5 can suppress the open source model.

GPT-4 Llama3 AI
Twice the delay! Intel disclosed the reasoning performance of Xeon 6 processor for Meta Llama 3 model

Recently, Meta launched its MetaLlama3 open source model with 8 billion and 70 billion parameters. The model introduces new functions such as improved reasoning and more model dimensions, and uses a new marker, aiming to improve the efficiency of coding language and improve the performance of the model. Even though the Llama3 model with 8 billion parameters has higher parameters than the Llama2 model with 7 billion parameters, when BF16 reasoning is run on the AWSm7i.metal-48xl instance, the reasoning delay of the overall prompt is almost the same.

Meta Llama three
Google released a new code model Code Gemma: small parameters, strong performance, low hardware requirements!

Google recently released a new code model CodeGemma for enterprises and developers. The model has small parameters, strong performance and low hardware requirements. CodeGemma has three versions: basic pre training, instruction fine-tuning and fast reasoning, which show excellent performance in benchmark tests. The above is related to the CodeGemma model released by Google. For more details, please refer to the open source address and technical report.

CodeGemma AI Headlines

Hot text

3 days
7 days

Business opportunity of webmaster

Business cooperation Infringement complaint Advertising services Copyright Notice recruit