They vibe-coded from the bottom — ForkLog

# Vibe coding from the bottom

A guide to launching open-source AI models from deep GitHub

In the development of AI, a vector has emerged in which decentralization and open-source code make it possible to go beyond popular commercial solutions. Local LLMs let you work with data privately, flexibly configure the system for your own tasks, and independently control the environment in which it’s used. At the same time, running such models requires understanding the basic tools—from repositories and model weights to cloud environments and technical specifications.

In the new piece, ForkLog will explain how to get started with autonomous AI models without spending money, what resources beginners should use, and what developers of OS solutions offer.

First impressions

For developers of open AI models, there are two main platforms—GitHub and Hugging Face. The first is traditionally used for publishing source code, documentation, and installation scripts, while the second has become a global hub for model weights, datasets, and ready-made ML solutions. On Hugging Face, hundreds of thousands of trained neural networks are published—ranging from tiny language models for smartphones, alternative generators of media content, to specialized algorithms for scientists and enthusiasts.

Choosing the necessary model is helped by community activity metrics. On GitHub, they are represented by the number of stars, the regularity of updates (commits), and the speed of resolving issues (issues).

It’s also important to check the origin of the product and the authenticity of the repository. Popular OS builds are regularly used as bait by cybercriminals who distribute malicious code under the guise of well-known AI tools.

The next stage of getting acquainted with local AI models is to test their functionality in practice. For users without powerful hardware, there are free and conditionally free cloud platforms

The most popular solution is Google Colab—an online environment that provides access to graphics processing units (GPU) directly from your browser. With a free subscription, you can work on a system with an Nvidia Tesla T4 accelerator for an average of two to four hours depending on the load. Alternatives include Kaggle Notebooks and Hugging Face Spaces. The latter lets you interact with models through ready-made web interfaces such as Gradio or Streamlit.

When working with federated solutions as well, you should consider the legal side. Many popular projects are available under classic licenses, such as MIT or Apache 2.0, which makes it possible to use them—including for commercial purposes—with minimal restrictions.

However, there are also specific approaches. Meta distributes its flagship models under its own Llama 3.1 Community License, which requires obtaining special permission if the service’s monthly audience exceeds 700 million users.

Strict copyleft licenses, such as the GNU General Public License, are also found—requiring the source code of all derivative products to be made public.

My personal ChatGPT analogue

From the enormous number of autonomous general-purpose LLMs (analogues of ChatGPT or Gemini), choosing the right model is helped by independent rankings based on blind testing and performance metrics such as Open LLM Leaderboard and Chatbot Arena.

Dashboard of open LLMs. Source: llm-stats. The gold standard of the segment is considered to be the Llama family of models by Meta and Qwen from Alibaba. These models work well with long context, handle multi-step requests, and are suitable for vibe coding and programming tasks. Thanks to the open framework Ollama, installing them is reduced to a single command.

During a test conducted while writing this material, the qwen3.5:2b model managed to run on a laptop without a discrete graphics card based on Core i7 with 8 GB RAM and an SSD, while closing heavy applications: messengers and browsers.

Source: Ollama. “2b” means 2 billion parameters. The higher the value, the more complex relationships the neural network can capture. For example, a 2b model will learn basic grammar and simple commands, while a 122b model will remember facts from quantum physics, the subtleties of legal documents, and learn to plan tasks ten steps ahead.

Each parameter takes up physical space on the hard drive and, most importantly, in RAM. The 2b model used about 4–5 GB of RAM and became the maximum model size for running on such a machine. At the same time, the response to the simplest request “hello!” was generated in almost three minutes.

Screenshot: ForkLog. Approximate model tiers:

• 0.5b-2b. Fast; can run on older laptops and smartphones. Ideal for simple tasks (routing commands, basic summarization, autocomplete of short code strings). Prone to hallucinations on complex requests;
• 3b-4b. A balance of speed and quality. Good for mobile devices, smart home, and automation tasks. For example, you can ask a chat-bot to dim the lights in the room, turn on the air conditioner, or raise the barrier;
• 7b-9b. Require about 6–8 GB of free RAM. Powerful models with context understanding and deep logic, suitable for programming and working with large texts.

In his recent study of vibe coding in Web3, Vladimir Sleeper found that on a machine at the level of a MacBook Air with 16 GB RAM, qwen2.5-coder:7b, qwen3:8b, llama3.2:3b, and deepseek-r1:8b will fit. More powerful models require investing in a powerful PC with high-end graphics cards or installing on rented servers.

Private data processing, 3D printing, and protecting the user

The ways to interact with open AI models depend on the user’s level of preparation and the available hardware. There are projects packaged into convenient installers (files with a .EXE extension) or mobile applications that work “out of the box.” Others are abandoned GitHub repositories where installation turns into a multi-hour struggle with conflicts from outdated libraries.

Applied AI models today are used not only for generating text. Even a superficial analysis of the ecosystem makes it possible to identify dozens of specialized tools for specific tasks

Working with video and 3D:

• CogVideoX. An open model from Zhipu AI for generating video based on a text description. Lets you create realistic short clips, has open weights, and can be deployed in environments such as Jupyter or Colab if you have enough video memory;
• DepthCrafter. A tool for extracting depth-of-field information from video. Useful for VFX and 3D modeling specialists. It allows creating high-precision depth maps for each frame of a dynamic scene;
• TRELLIS (Morfx 3D). An advanced 3D asset generation system. The project allows creating high-quality three-dimensional models from images or text prompts, optimizing them for use in game engines.

Turning a photo of a train into an object for processing and 3D printing using the web version of the Morfx 3D model. Screenshot: ForkLog.Sound and recognition:

• CosyVoice. A multilingual speech synthesis model that supports voice cloning. Allows generating a realistic audio track while preserving the intonation and emotional tone of the original speaker;
• Whisper-WebGPU. An implementation of OpenAI’s speech recognition model rewritten to run directly in the browser using the WebGPU API. This means that audio transcription happens locally, ensuring full privacy without sending audio files to third-party servers;
• BirdNET-Analyzer. A neural network from Cornell University for determining bird species from their calls. Unlike the popular Merlin Bird ID app, which relies largely on cloud processing for some functions, BirdNET-Analyzer provides full control over the analysis process locally and can be used for bulk processing of gigabytes of field recordings.

Source: BirdNET.Programming and user protection:

• Screenshot-to-Code. A utility for converting a screenshot of a web page or a mobile app into clean HTML, Tailwind, or React code. Although the project often works together with paid APIs (Claude, GPT-4), the architecture makes it possible to connect open multimodal models;
• MinerU/Magic-PDF. A project for accurately extracting structured data from PDF documents. The model recognizes text, mathematical formulas, and tables, converting complex layout into Markdown;
• Fawkes. Makes changes to images that are invisible to the human eye, interfering with face recognition systems’ ability to identify a person. Loaded locally on a PC via a file with a .EXE extension and can be used for avatars on social media;
• Nightshade. “Poisoning” image pixels to confuse AI companies’ training algorithms if they do so without permission. For example, for the prompt “dog,” the model will generate an image of a cat.

Portrait of U.S. President Donald Trump before using Fawkes. Source: Library of Congress. After processing by Fawkes algorithms. Screenshot: ForkLog.

Fighting with libraries and the first success

After installing AI models with understandable UI/UX, it was necessary to figure out how easy it would be to deploy a heavy repository in the cloud—preferably for free.

FLUX.1 from the startup Black Forest Labs is one of the advanced image generation models, competing with corporate Midjourney and Nano Banana. With the required equipment, the software can run autonomously without internet access and allows bypassing censorship.

The test used the lightest free version, FLUX.1 Schnell. For convenience of interacting with open solutions, developers create target frameworks such as Ollama. For image generation, graphical interfaces like ComfyUI and Forge are popular.

During attempts to install the Forge implementation—cagliostro-forge-colab—an entire GPU access session from Google Colab had to be spent. The problem turned out to be a classic beginner’s mistake—mismatched versions of Python, the cloud environment, and the model itself. After four hours of vibe coding with the free version of Gemini 3 Flash, success could not be achieved.

In the end, I had to give up on installing the framework and move directly to deploying FLUX.1—but in the next free session, on another day

In practice, it’s more convenient to use free Google Colab on weekends: at that time, the platform often provides longer access

The model took about 34 GB of disk space on the cloud SSD. But all the processes associated with installation eventually used about 86 GB.

Used resources of the Google Colab cloud machine. Screenshot: ForkLog. At the first stage, the FLUX.1 Schnell model didn’t have enough video memory of the Nvidia Tesla T4 accelerator. The unadapted configuration hit the GPU limits until, after a series of simple code experiments, Gemini 3 Flash helped make adjustments by using step-by-step loading and memory cleanup. As a result, of the available 16 GB of video memory, about 3 GB was used during generation.

Screenshot: ForkLog. The process of creating one image took about seven minutes. Considering this is the free version of an open model, the result was pleasantly surprising.

Generated image created with FLUX.1 Schnell. Source: ForkLog. When trying several times to generate an image of rock singer Marilyn Manson in a Victorian style with a companion, FLUX.1 Schnell likely didn’t recognize the reference to a specific person and reproduced only a generalized visual template.

Generated image of the performer based on the prompt “draw Marilyn Manson in Victorian style” using FLUX.1 Schnell. Source: ForkLog.## Complex and incredible

Open neural networks have long been used not only for generating text and images, but also for narrower and more unusual tasks. A vivid example of a non-standard use of AI architecture is the GameNGen model, capable of recreating the gameplay of the classic shooter DOOM in real time.

Source: GameNGen/Github GameNGen doesn’t simulate the game in the usual sense—it sequentially generates video: the model predicts what the next frame should look like after a user action (for example, movement or a shot). Because of this, enemies, objects, and scene changes aren’t “calculated” by an engine; instead, they are visually reproduced as the most probable outcome.

Among autonomous systems, the Voyager project stands out—an AI agent for Minecraft. It independently explores the game world, gathers resources, and continuously self-trains.

The scientific community is also actively adapting open AI for its own needs—for example, using algorithms to decode history. Thus, researchers from Tel Aviv and Munich universities trained the Akkademia model to directly translate ancient Akkadian cuneiform into English. It makes it possible to process thousands of damaged clay tablets, accelerating archaeologists’ work by dozens of times.

No less interesting is the MinD-Vis project. This system analyzes functional MRI data and tries to reconstruct images that the subject is observing at the moment of scanning. In other words, it generates an interpretation of what a person saw based on patterns of brain activity.

Initiatives like these prove that artificial intelligence has turned into a universal tool for understanding and modeling reality. Shifting the initiative from closed corporate APIs to open-source code creates an entirely new technology development paradigm. Today, any researcher, developer, or enthusiast can deploy infrastructure that just a few years ago required multi-million investments in server farms.

The development of the ecosystem inevitably comes with improvements to the user experience: complex scripts are being replaced by intuitive interfaces and automated deployment environments. Using tools like Ollama and Forge demonstrates that privacy, the absence of censorship, and high performance can coexist harmoniously in a single software solution. The future of the AI industry today depends largely on how strong, scalable, and independent the open ecosystem will remain.