I. Introduction
Hello everyone, Brother Yong is here to share the AI model again. The Xunhuo Large Model 2.0 released by Xunfei a few days ago has quickly entered our circle. In order to share more and better models with everyone, it is necessary to share the Xinghuo Large Model 2.0 I have done it, I will pass it on, and I hope everyone will also pass it on.
I have been using it for a few days, so I have summarized some of the main advantages of iFLYTEK's artificial intelligence model Xinghuo model 2.0:
1. Obtained a number of technical patents: HKUST Xunfei has ranked first in AI patent applications for several times, and the continuous research and development and output of technology have provided strong support for the continuous evolution of the Xinghuo large model.
2. Accumulation of cutting-edge technologies: HKUST Xunfei has been cultivating Chinese pre-training models for many years. Xinghuo Large Model 2.0 is Xunfei’s first 100-billion-level deep knowledge enhancement large model, which can support multi-language access, multi-task fusion and multi-mode State integration, and continue to introduce cutting-edge technologies in the industry, and continue to explore and develop AI cutting-edge technologies.
3. Powerful knowledge enhancement capability: iFLYTEK's artificial intelligence large model is based on a general-purpose deep learning model, which combines knowledge enhancement, multi-task training and cross-modal fusion technologies to better understand and process natural language, Including speech recognition, text generation, image recognition and other tasks.
4. Rich application scenarios: iFLYTEK's artificial intelligence model has been applied in many fields, such as intelligent voice, intelligent customer service, intelligent education, intelligent medical care, etc., and has been used in traditional industries such as industry, agriculture, energy, and environmental protection. industry has been applied.
The above 4 points are my conclusions. Of course, with the development of science and technology and the progress of human beings, there will be 3.0/4.0...many versions of the Spark Model. I hope it will bring us more surprises.
2. Authoritative data evaluation results
The following information comes from [Economic Information Daily], specifically from the "Artificial Intelligence Large-scale Model Experience Report 2.0" in August 2023. You can check it out. The data is accurate and I have not made any changes.
Original link: https://xhpfmapi.xinhuaxmt.com/vh512/share/11634934?channel=weixin
You can see in the picture above that Xunfei Xinghuo is only one point less than the answer of humans, and the [ tool efficiency ] ability crushes the audience, and super VIP must be firmly held.
2.1 Description of Tool Efficiency Improvement Indicators
The work efficiency improvement inspection of large models usually includes two aspects: tools and innovations.
Tool efficiency refers to whether the large model can provide effective tools to improve work efficiency. These tools can
Including automatic code generation, automatic data analysis and visualization tools, natural language processing, file organization,
Key content summary integration and machine translation tools, etc. Large models with these tools can help people complete work tasks faster and improve work efficiency, and it is also one of the first directions for large models to land.
The aspect of generative innovation refers to whether the large model can provide new ideas and methods to promote the improvement and upgrading of working methods. These innovations can include new business processes, innovations in topic selection, innovations in content, etc. AIGC
Capabilities can help people break out of traditional ways of working and boost work efficiency.
Weight ratio: tool efficiency improvement (50%), generation innovation (50%)
3. Multimodal function
Separate explanation 【Multimodal function】
Multimodal capability refers to the ability to process and understand multi-source modal information, including different types of data such as text, image, video, and audio. Multimodal machine learning aims to achieve the ability to process and understand multi-source modal information through machine learning methods. At present, the popular research direction is multi-modal learning among image, video, audio and semantics.
[Xunfei Spark Cognitive Model] has 7 core capabilities, namely, text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, coding ability, and multimodal ability. This is also a feature we need to test in a moment.
It has cross-domain knowledge and language understanding capabilities, and can understand and perform tasks based on natural dialogue. Continue to evolve from massive data and large-scale knowledge, and realize the closed loop of the whole process from proposal, planning to problem solving.
4. Mars Large Model 2.0 Test
I have explained the advantages and rankings of the Mars Large Model 2.0 to everyone, so let's test it in detail next.
test steps
Official website address: After registration, click [Enter Experience].
We can see that there is a Spark Assistant, and there are a lot of prompts in this place to help us quickly locate the problem we want to solve.
4.1 Mathematics ability
Mathematics can be said to be the starting point of everything in our science and engineering, so let's test this first.
Test content: [Calculus reciprocal problem, medium difficulty, and complete steps to solve the problem. 】
We can see the solution steps of the whole problem, which is very clear. The method of calculating [π] is also good.
4.2 Tanabata test
The test sentence is relatively difficult, and I added a lot of modifiers.
[What is the best gift for Qixi Festival this year? The other party is a woman, 25 years old, gentle and coquettish, with an independent look, a professional woman. 】
This one is quite difficult, and there is no single answer, but 10 help options are given. At the beginning, these answers are similar to our own answers, which is not bad. This yoga course is unexpected.
4.3 Image generation
Test text [Help me draw a real photo, there is Feixue, there is a girl wearing open-toed gloves, short light yellow hair, bright eyes, black pupils, cherry mouth, more delicate, high-definition pictures, pink sweater with a hat , exhale into the hand, and look up at a 35-degree angle. ], let's compare them in turn.
1. Real person (conform)
2. Flying snow (conforming)
3. There are girls who wear open-toed gloves (70%)
4. Light yellow hair (conform)
5. Bright eyes, black pupils, cherry mouth (conform)
6. Pink sweater with hood (conforms)
7. Exhale into the hands (does not meet)
8. Profile photo with head raised at 35 degrees (80%)
The overall comparison is still very good. We can also generate more times, I framed the button on the way, just find a suitable picture and leave it.
4.4 Speech Generation
It can directly generate speech, which is much better than other platforms, so I won’t say who to compare it with, let’s try it out and it will be clear.
Let's take a difficult test content [Generate a girl clip voice to celebrate the birthday of the 175-year-old grandfather. 】
You can listen to it yourself, it's very interesting.
4.5 Video generation
Test text【Generate a 15-second slow three dance video. 】
Click to open the video effect as follows:
After explaining the slow three dance, the AI narrator is also very wow.
4.6 PPT generation
We can [select plug-in] when we enter, I will test it by generating PPT here.
Test text [Generate a Linux-based file operation PPT, the number of pages is within 15 pages. 】
Download to see the effect:
The generation effect is still very good, and the whole content is very complete.
4.7 Code generation
In order to make the test more difficult, I chose to test assembly language.
The generated code, even if you don’t know assembly language, at least know that [;] is an assembly comment, so the following code is fine, it must be assembly.
Bash
section .data
result db 1
section .text
global _start
_start:
; 调用递归函数
mov eax, 5
call factorial
; 退出程序
mov eax, 1
xor ebx, ebx
int 0x80
; 递归函数
factorial:
cmp eax, 1
jle end_recursion
dec eax
push eax
call factorial
pop eax
add esp, 4
imul dword [result], eax
end_recursion:
ret
4.8 Fiction Assistant
There are still many types of novel assistants.
Now when we read novels, most of the time we read cool articles, that is, our favorite type of articles, so when we generate them, we can generate them according to our thoughts at the time, let’s try.
Test content: [Generate a novel about Wu Song who travels through the Outlaws of the Marsh and escapes with a one-on-one fight against Fang La. 】
A small article of about 600 words was generated. I read the whole article. It was written from the perspective of God. I understood the meaning of the requirements, and it was more vivid and powerful. The sentences used were relatively simple. The complex sentence structure clearly described what I expected.
V. Summary
I have passed multiple content tests here, and we can see that the effect of the feedback content is very reliable, and pictures, voices, and videos can be directly generated, and PPT and resumes can also be generated. These functions are really good in other places. It's hard to find, and it may take multiple applications to complete, but here it is all solved in one place, and the tips given in the [Help Center] are really considerate.
On the whole, Xinghuo Large Model 2.0 is very practical and useful, and it is worth trying and using by users. Everyone recommends it to each other. I believe that such an easy-to-use tool will surely [Xinghuo Liaoyuan].
Experience address 
https://xinghuo.xfyun.cn/?ch=bl_9xuGqP
The review of registration through this experience address is faster (the registration application is available and passed in seconds), and it is strongly recommended to apply through the exclusive link;
Developer rights and benefits: This experience address application can also have a higher Spark large model API test quota, which is 30% more than ordinary channel applications, and can apply for up to 500w Tokens.
Come and experience it.