A new free programming tool! Two times faster than Copilot and 20% more accurate | Released by Feishen Technology

Jin Lei comes from Ao Fei Temple
Qubit | Public account QbitAI

A domestic programming artifact that instantly beats Microsoft GitHub Copilot.

How fast?

While twice as fast as GitHub Copilot , the accuracy is also improved by about 20% .

Not only that, while being "fast" and "accurate", its functions are also relatively comprehensive, including:

Automatic code completion, natural language code generation, automatic addition of comments, intelligent bug finding, code interpretation, automatic generation of unit tests, etc.

Highlights: Supports 80 languages , completely free of charge !

At present, Fitten Code has become the first place in the VSCode plug-in market trend list !

This is the newly launched Fitten Code from Feishen Technology , which is completely developed based on the domestic deep learning framework Jittor and the large code model developed by Feishen.

It is worth mentioning that its core development team all have Ph.D. degrees from Tsinghua University; and while they were in school, they were the main developers who promoted the open source work of Jitu. After graduation, they established Feishen Technology Ventures and continued Promote the development of graphics and develop the JNeRF neural rendering library and JittorLLMs large model reasoning library as well as the Fitten Code AI programming assistant released this time.

So next, let’s take a look at the actual measurement results of Fitten Code.

Faster and more accurate than GitHub Copilot

First of all, we still put Fitten Code and GitHub Copilot together to compete on the same stage in terms of speed .

Also faced with the task of "writing a ResNet model", the difference in speed is visible to the naked eye - the average delay of Fitten Code is only 300ms , while the first delay of GitHub Copilot is as long as a full 5 seconds.

Not only that, there is also a big gap in the quality of the generated code results.

Fitten Code generates the complete code of ResNet, and it can be completed interactively; but on the GitHub Copilot side, the generated code contains a large number of repeated code fragments.

Low latency is certainly an important aspect of a programming assistant, but accuracy can be said to be even more valuable.

Judging from the results in the HumanEval test set, Fitten Code has achieved "the best of both worlds":

Fitten Code's Pass@1 accuracy reached 60.1% , a significant increase compared to Copilot's 49.5%.

Fitten Code also shows its efficient side when it comes to complex algorithm tasks. For example, we issued a task like this:

Please use Python to implement the algorithm for the longest ascending subsequence, which requires a time complexity of O(nlogn).

As can be seen from the results, Fitten Code completed this complex task very accurately.

Let's take a look at GitHub Copilot. It can only implement a non-optimal algorithm of O(n^2).

Not only that, Fitten Code will complete a large amount of code at once when conditions permit.

Unlike other products that only complete 1-2 lines at a time on average, Fitten Code completes 3-5 lines on average, which greatly improves the completion efficiency.

Therefore, after several rounds of "competition", it is not difficult to find that Fitten Code beats GitHub Copilot in terms of corresponding speed, code completion amount and accuracy.

More function display

As we just mentioned, Fitten Code currently supports over 80 programming languages, including:

Python, Javascript, Typescript, Java, C, C++, Kotlin, PHP, Ruby, etc., and supports Visual Studio Code.

So next, let's take a look at the performance of this large model-driven code generation tool in real scenarios.

Automatic code completion

Fitten Code can automatically supplement the missing parts of the code. This intelligent experience makes typing code faster.

Natural language generated code

Fitten Code can achieve semantic-level translation of code and support mutual translation between multiple programming languages.

Just use comments (#) or dialogue to describe the functions that the code needs to implement, and Fitten Code can automatically generate code that meets the comment requirements, greatly reducing the time and energy of manual writing.

In addition, using comments as a guide, the code generation plug-in can quickly generate project structure, function skeleton, interface call and other code snippets to help quickly build the project.

Automatically add comments

As for writing comments, with Fitten Code you can say goodbye to manual work.

It can automatically generate relevant comments based on the code and provide clear and easy-to-understand explanations and documentation by analyzing the logic and structure of the code.

In addition to the above functions, Fitten Code also provides a wealth of practical functions, such as intelligent bug finding based on the selected code.

It also has the function of automatically generating unit tests, which can automatically generate corresponding test cases based on the code to improve code quality and reliability.

Not only that, Fitten Code also has the ability to explain the meaning of the code, helping users understand the logic and structure of the code more deeply.

All in all, this domestic programming artifact is designed to make programming more efficient from all aspects.

how to use?

Fitten Code is not only free and full-featured, but its installation method is also extremely simple.

Taking Visual Studio Code as an example, search for "Fitten Code" in the extension page:

Then click "Install":

Finally, just register and log in to use it~

Super simple, no problem!

In addition, Fitten Code also supports JetBrains series IDEs such as IntelliJ IDEA and PyCharm.

About the team

Fitten Code was developed by Beijing Feishen Technology Co., Ltd. The core team all graduated with Ph.D. degrees from Tsinghua University. As the main developer, they have open sourced the Jittor deep learning framework, which is one of the mainstream deep learning frameworks in China, and participated in the open source JNeRF The neural rendering library and JittorLLMs large model inference library have received a lot of praise from the industry.

Team members have outstanding technical abilities and have won gold medals in international supercomputing competitions, Informatics Olympiads, and ACM gold medals, and have published papers in top international conferences and journals on computer graphics, computer vision, and artificial intelligence such as CVPR, SIGGRAPH, TOG, TIP, and CVM. Published many articles, it is a top multi-disciplinary team in the international and local professional fields with artificial intelligence, deep learning, high-performance computing, system design, hardware architecture and other interdisciplinary teams.

Are you excited about such a free and easy-to-use domestic programming artifact?

Click here ???? Follow me and remember to star~