EHCode Hub compares GLM-5.1, DeepSeek V4 Pro and MiMo V2.5 Pro on 20 coding tasks, with automatic project launching, a smart code editor and an embedded opencode assistant.
An academic project carried out by a team of students with technical assistance from Techno Enjaz to build a platform called EHCode Hub, which works as a benchmark harness for comparing AI code-generation models and also includes an intelligent coding assistant and a code editor that analyzes, fixes, and converts code between languages. The platform compared three models: GLM-5.1 by Zhipu AI, DeepSeek V4 Pro, and MiMo V2.5 Pro by Xiaomi, across 20 unified programming tasks in ten development categories.
The platform's server was built with Node.js using only its built-in modules, with no external dependencies, alongside an Arabic interface in HTML5, CSS3, and vanilla JavaScript, and the EHCode assistant, built on the open-source opencode agent, is embedded through an iframe. The three models produced 60 projects, and GLM-5.1 topped the evaluation with a weighted score of 8.78 out of 10, followed by DeepSeek V4 Pro at 8.75 and MiMo V2.5 Pro at 7.96.
| Item | Details |
|---|---|
| Project type | Code-generation model evaluation platform with an intelligent coding assistant |
| Field | Software engineering, large language models (LLMs), developer tools |
| Project status | Platform running locally on Windows, tested on 60 generated projects |
| Techno Enjaz role | Technical support and assistance to the students during project development |
| Core technologies | Node.js (built-in modules), HTML5, CSS3, Vanilla JS, opencode, MCP, Playwright, Batch Scripts |
| Outputs | Interactive project viewer, automatic launching, per-project guide, model comparison table, smart code editor, EHCode assistant with web and CLI interfaces |
As large language models spread into code writing, developers and organizations struggle to choose the right model, because common evaluations are either narrow in scope or rely on automated metrics that do not reflect the real quality of output in actual projects. Models may also excel at frontends while falling short on backends or mobile apps.
The project adds practical problems: running generated projects by hand wastes time installing dependencies and freeing ports, many of these projects have no documentation, and current evaluation environments offer no coding assistant to help understand the code during evaluation. This led to the idea of a single platform that combines automatic scanning, display, launching, documentation, comparison, and coding assistance.
The project was carried out academically by the students, with technical assistance and support from Techno Enjaz during development. It is presented among the office's work as an example of its support for developer-tool projects and applications built on large language models.
/models folder and classifies each project by its files: package.json for Node.js, requirements.txt for Python, pubspec.yaml for Flutter, and anything else as a static project./api/launch and stops any other project to free the port.npm install then node, Python projects run pip install then python, and for Flutter projects the user is directed to run them manually because of their complex environment./api/probe every 1.5 seconds for up to 20 attempts, then loads the project in an iframe or shows the reason for failure.| Component | Function |
|---|---|
| Viewer server (port 8800) | Scans and launches projects and serves the interface, using only the http, fs, path, url, and child_process modules |
| EHCode assistant (port 8801) | Chat interface built on opencode web, embedded via iframe, with model selection and session management |
| Smart code editor | Solve, Fix, Convert, Ask, and code analysis buttons, with syntax highlighting and line numbers |
| PHP-to-Laravel converter | Restructures a PHP project into Laravel's structure (Routes, Controllers, Models, Migrations, Blade), with both versions runnable and a live output panel |
| Model comparison page | Indicator table and weighted quality-criteria table, with the winning model announced |
| Unified launcher start.bat | Checks for node and opencode, frees ports 8800 and 8801, starts the viewer and the assistant, then opens the browser; stop.bat shuts them down |
| Portable EHCode distribution | A standalone ehcode.exe that needs no Node.js, with install.bat enabling Playwright MCP |
The assistant is available through a web interface and a command-line interface (CLI). In the code editor, the book shows an example of converting code from Python to C++, and an analysis report that explains what the code does and suggests improvements to quality and maintainability.
Each model produced 20 projects, and start.bat run files were present and error-free in all projects (20/20 for each model):
| Indicator | GLM-5.1 | DeepSeek V4 Pro | MiMo V2.5 Pro |
|---|---|---|---|
| Node.js projects | 3 | 3 | 4 |
| Python projects | 2 | 0 | 2 |
| Static projects | 13 | 17 | 12 |
| Flutter projects | 2 | 0 | 2 |
| Projects with a README file | 5 | 1 | 2 |
A key observation is that DeepSeek V4 Pro produced HTML pages instead of Flutter apps for the two mobile tasks, while GLM-5.1 and MiMo V2.5 Pro produced real Flutter apps as required.
The models were scored on eight weighted criteria: functional quality, performance, reliability, security, usability, maintainability, cost efficiency, and portability:
| Model | Top scores (out of 10) | Weighted total |
|---|---|---|
| GLM-5.1 | Functional quality 9.2, usability 9.0, maintainability 9.2 | 8.78 |
| DeepSeek V4 Pro | Performance 9.2, reliability 9.0, security 8.8, cost efficiency 9.5 | 8.75 |
| MiMo V2.5 Pro | Portability 9.0, with balanced performance on the other criteria | 7.96 |
On this basis, the project recommended GLM-5.1 for general use, DeepSeek V4 Pro for backend-focused projects, and MiMo V2.5 Pro for frontend projects. The platform also proved able to manage processes, free ports, and launch projects automatically; documented examples include running a blog platform and a 3D Earth globe from GLM-5.1's projects.
According to the directions set out in the project, the platform could be developed by:
If you are working on a smart tool for source-code analysis and need technical help developing the idea or building a prototype, you can contact the Techno Enjaz office to discuss your project's requirements and the appropriate scope of support.
It is a platform built by a team of students with technical assistance from Techno Enjaz that works as a benchmark harness for comparing AI code-generation models. It scans and automatically launches generated projects inside an iframe, provides a guide per project and a comparison table, and includes a smart code editor and a coding assistant called EHCode, built on opencode.
The Node.js server classifies each project by its files (package.json, requirements.txt, or pubspec.yaml), then runs npm install or pip install and starts the server in the background after freeing the port, or serves static projects directly. The frontend polls readiness every 1.5 seconds for up to 20 attempts before showing the project, while Flutter projects are run manually.
The server was built with Node.js using only its built-in modules (http, fs, path, url, and child_process) with no external dependencies, and the interface with HTML5, CSS3, and vanilla JavaScript. The assistant is based on opencode with MCP and Playwright support, and the components are started with Batch files on Windows.
GLM-5.1, DeepSeek V4 Pro, and MiMo V2.5 Pro each produced twenty projects. GLM-5.1 achieved the highest weighted score (8.78 out of 10), followed by DeepSeek V4 Pro (8.75) and MiMo V2.5 Pro (7.96). DeepSeek produced HTML pages instead of Flutter apps for the two mobile tasks, and GLM-5.1 had the best documentation with five README files.
The evaluation relied on 20 tasks and only three models, and the project states that score registration and updating the results data are still needed to complete the quantitative evaluation, without detailing how scores were assigned. The platform currently runs on Windows; proposed developments include automating evaluation and integrating HumanEval, MBPP, and SWE-bench.
Reach out to Techno Enjaz engineering team to discuss technical feasibility, prototype hardware, and development scope.
Discover more engineering prototypes and academic implementations
A Laravel web platform for publishing official news and receiving user reports on suspicious content, reviewed by editors who publish corrections linked to the fake item.
A Laravel and Bootstrap web platform for a digital voting cycle with three roles: voter registration, admin-reviewed candidacy requests, and one-time confirmed voting.
A Flutter app with a Laravel back end that matches a student's Project-DNA skills profile to projects and generates a work plan, risks, and an advisory readiness score.