How do I convert a GitHub repo to a text file?

Paste the repository URL above and the file tree appears as a checkbox list. Untick what the model does not need, generate, and you get one text file: directory structure at the top, then each selected file under a header with its path. Nothing to clone, nothing to install, free and no sign-up.

How do I convert a private GitHub repository?

Paste a personal access token alongside the URL. Create a fine-grained token scoped to just that one repository with read-only Contents permission and a short expiry — that is the least privilege the API will accept for this. The token is used to call GitHub from your browser and is not stored on our side.

How many tokens is my codebase, and will it fit?

The counter updates as you tick files, so you can see the number before you paste anything. As a rough guide, source code runs around 3 to 4 characters per token, so a 500 KB selection lands near 140,000 tokens. That fits a 200K context window and will not fit into a single chat message on most interfaces.

Which files should I exclude before feeding a repo to an LLM?

node_modules, vendor, and any dependency directory first — they are usually the overwhelming majority of the bytes and contribute nothing about your code. Then lockfiles, build output, minified bundles, test fixtures, and binary assets. On a typical JavaScript project that pass alone often removes over ninety percent of the tokens.

How does this compare to gitingest, repomix, or code2prompt?

Those are strong tools and the honest difference is workflow, not capability. They are command-line first, scriptable, and the right answer inside CI or a batch job. This runs in a browser tab with a checkbox tree, so you can see the repo, deselect what you do not want, and watch the token count move — better for the one-off case where you are still deciding what to include.

Is it safe to put my code through this?

Public repositories are read through the GitHub API from your browser, so the code goes from GitHub to you. For anything confidential the stronger option is the local folder converter, which never uploads at all. Either way, scan your selection for API keys and credentials before pasting it into a model — that is the step people skip.

BAGEL by Bytedance: The Open-Source Multimodal AI Alternative to GPT-4o

ByteDance’s BAGEL: The Open-Source Multimodal AI Powerhouse

ByteDance just shook up the AI world by open-sourcing BAGEL , a unified multimodal foundation model that “reads, sees, generates, and reasons” across text, images, and even video. With 7 billion active (14B total) parameters and a clever Mixture-of-Transformer-Experts (MoT) design, BAGEL is built to rival closed-source giants like OpenAI’s GPT-4o or Google’s Gemini – while remaining free to inspect and modify. In plain terms, BAGEL is an all-in-one AI: it can understand what it reads and sees, generate new content in multiple media, and even perform complex vision-language reasoning. Unlike proprietary models that hide behind APIs, BAGEL’s Apache 2.0 license means researchers and developers can use, fine-tune, and study it without restriction.

In practice, that means a single model handles tasks that used to require separate systems. Rather than juggling a text-only chatbot, an image generator, and a video predictor, you can feed BAGEL a prompt like “ A steampunk cat playing the violin ” or an actual photo, and get back images, captions, or even predictions of what happens next in a scene. It supports text-to-image generation, image editing (including style transfer and scene outpainting), video generation/prediction, and more . In short, BAGEL can do things like:

Vision-Language Understanding: Answer questions about an image (e.g. “How many cars are in this street photo?”), caption scenes, and analyze objects. It tops benchmarks like MMBench and MM-Vet, outperforming other open models (e.g. Qwen-VL, InternVL).
Text-to-Image Generation: Create vivid, high-quality images from text prompts. On generation tests (GenEval), BAGEL scores around 0.88 (higher than many peers), producing images that not only look real but also faithfully match the prompt and style.
Image Editing (Free-form & Stylization): Take an existing photo and transform it. You can tell BAGEL to “ change the sky to a stormy night ” or “ make the dog wear a wizard hat ,” and it will do so with high fidelity. It even supports multi-step (chain-of-thought) editing : you can ask it to plan and apply a series of edits, and it will reason through them step by step. Unlike some models that only change fixed parts, BAGEL can add/remove elements, transfer artistic styles between images, or outpaint beyond the original borders for seamless scene expansion.
Video Understanding & Generation: Because it was trained on interleaved video data, BAGEL can predict future frames or generate short clips. Give it one frame of a video, and it can infer what comes next or fill in missing parts. For example, from a photo of a street it might infer what a nearby alley looks like or what a person sees if they “walk forward”. This shows BAGEL’s world-modeling ability – an internal sense of 3D space, object movement, and physical consistency.
3D & Multiview Reasoning: BAGEL can imagine scenes from different angles. Given one view of a room or object, it can synthesize plausible side or back views (multiview synthesis) and even simulate 3D rotations. This means it understands depth and spatial layout, so you could ask it to rotate a photo of a car or render the back of a building it’s “seen.”

These are not hand-coded tricks – they emerge from scale and training . By mixing text, images, and videos together in training, BAGEL learns how language and vision relate in the real world. The ByteDance-Seed team fed it trillions of tokens from a blend of sources (web text, labeled images, video frames, etc.). Importantly, the data streams were interleaved: a training sequence might include some text, then an image, then more text, then a video frame, all mixed together. This forced the model to build unified representations – for instance, understanding how a caption describes its image or how one video frame follows another. BAGEL even predicts entire groups of tokens at once (like whole sentences or images) for efficiency, rather than a single word/token. All this heavy lifting results in “emergent” capabilities: the model wasn’t explicitly taught to do 3D, outpaint, or chain-of-thought, yet it can reason in these ways.

Under the hood, BAGEL uses a Mixture-of-Transformer-Experts architecture. In simple terms, it’s a big Transformer with two “expert” decoder stacks: one geared toward understanding (comprehension) and one for generation (creative output). Both experts share the same attention context, so they stay in sync. Visual inputs go through dual encoders : a Vision Transformer (SigLIP-L) captures high-level meaning from images, while a VAE-style encoder from the FLUX.1 project handles pixel-level details. Text tokens are processed autoregressively (left-to-right), whereas vision tokens use bidirectional attention so they can be interpreted like a whole image at once. This hybrid scheme lets BAGEL answer complex questions about pictures, write captions, or generate photorealistic scenes – all with one model.

Key Features & Capabilities

BAGEL packs an impressive feature set for image/video tasks. Here are some highlights:

Image Reasoning & Vision-Language Tasks: BAGEL excels at questions and analysis about images. Ask it “ How many people are playing soccer in this photo? ” or “ Describe this scene in one sentence ,” and it gives accurate answers and captions. In head-to-head tests, it outperforms similar-sized open models: ByteDance claims it beats Qwen2.5-VL-7B on image understanding benchmarks. In vision benchmarks like MMBench and InternVet, BAGEL sits at the top, showing very strong object recognition and scene interpretation.
Text-to-Image Generation: Prompt BAGEL with a description (“ A robotic owl perched on a futuristic tree branch ”) and it will conjure a stunning image. In experiments (GenEval), BAGEL achieved a high relevance score (≈0.88), surpassing many popular models like Stable Diffusion 3 (SD3-Medium) which scored around 0.74. Notably, its images aren’t just realistic pixels: they also tend to adhere very well to the prompt’s details and style. In practice, this means you could use BAGEL to generate illustrations, concept art, or any visual content from natural language prompts.
Advanced Image Editing (Style Transfer, Outpainting, etc.): One of BAGEL’s standout strengths is free-form image editing . You can give it a photo or painting and instruct pretty much anything. For example, you might say “ make the sky a stormy night with lightning ” or “ turn the dog blue and swap it with the cat ,” and BAGEL will carry out the edit seamlessly. It can add or remove elements (“ take out the car on the left ”), change expressions or weather, or overlay new objects, often more accurately than other open-source systems. BAGEL even supports style transfer : you could ask it to re-render an image in the style of Van Gogh or a particular movie, and it will do so while preserving content. Similarly, outpainting (extending an image beyond its original borders) comes naturally: the model knows the “world” beyond the frame. In benchmarks like GEdit-Bench (image editing), BAGEL scores much higher than competitors – reportedly around 7.36 out of 10.
Chain-of-Thought & Multi-Step Edits: For very complex edits, you can prompt BAGEL to “think in steps.” For instance, you might say “First plan the steps to make this city scene look futuristic, then apply them.” BAGEL can literally break the task into a logical sequence and execute it. This emergent “chain-of-thought” editing means it can handle multi-turn, multi-step instructions involving images.
Video Frame Prediction & Generation: Because BAGEL was trained on video clips, it has a basic sense of motion and temporal consistency. Give it a single frame, and it can predict what the next frame or two should look like. For example, showing it a photo of someone walking, you could ask “ What will the next moment look like? ” and it will generate a plausible continuation. This capability extends to “ future frame prediction ” – like simulating how a scene will evolve. You can also use this for video generation: provide a short text prompt or an initial frame, and BAGEL can create a small video sequence (or series of frames) that follow the logic.
3D & Spatial Manipulation: BAGEL’s understanding of space is remarkable for a 2D model. It can effectively perform 3D manipulation tasks. For instance, given one view of a room, it can generate a realistic image of what the unseen side might look like (multiview synthesis). You can also rotate objects within an image: if you have a photo of a car from the front, you could ask for a 3D rotation or a back view, and BAGEL will produce it consistently. This “spatial reasoning” comes from its training on diverse data – it has learned about object shapes, perspectives, and depth.
World Modeling & Navigation: Beyond static images, BAGEL has an emergent sense of “world context.” It can handle prompts like “If I walk forward in this scene, what will I see next?” or “Imagine this city scene at midnight.” These tasks show that BAGEL is building an internal world model: it understands relationships between objects, lighting, and even physical factors like wind or rain. It’s essentially doing multi-modal reasoning , combining text and visual cues to imagine changes or movements in the scene.

Benchmarks & Performance: In testing, BAGEL impressively outperforms other open models across the board. ByteDance reports it beats Qwen2.5-VL-7B on vision benchmarks and surpasses Janus-Pro-7B and Flux-1-dev on image generation tests. It even bests Google’s Gemini-2 experimental model on editing tasks (GEdit-Bench). And thanks to its open nature, it achieves this without the usual paywall – researchers have released checkpoints, code, and even a demo space.

All of this adds up to a model that can handle “multimodal understanding and generation” natively . In fact, the authors wrote that scaling a decoder-only model with such diverse data enabled “advanced multimodal reasoning abilities such as free-form image manipulation, future frame prediction, 3D manipulation, and world navigation” . Those are exactly the tricks BAGEL can do.

Inside BAGEL’s Architecture & Training

It helps to know how BAGEL is built. The core is a decoder-only Transformer with a Mixture-of-Experts setup. Essentially, there are two Transformer “experts” running side-by-side: one fine-tuned for understanding input, the other for generating output. Both experts see the same input sequence (text, image tokens, etc.) but specialize differently. Critically, they share the attention context, so the understanding and generation sides stay coherent. This design boosts capacity without needing two separate models.

Dual Vision Encoders: When an image comes in, BAGEL processes it through two encoders. First, a Vision Transformer (SigLIP-L) extracts semantic features – the big-picture stuff (objects, scene category, etc.). Second, a VAE-based encoder (from the FLUX.1 project) captures pixel-level details and textures. By combining these, BAGEL has both high-level understanding and low-level detail. For videos, frames are treated similarly, and a special “spatial encoder” can fuse multiple frames if needed (more details are in the paper).

Tokenization & Attention: Text tokens follow a causal (left-to-right) scheme, like in GPT, while vision tokens use bidirectional attention. This means BAGEL can both predict sequences (for generation) and fill in context (for understanding an image as a whole). It also employs a trick called “next group prediction”: instead of just predicting the next word, it might predict a whole chunk (like a sentence or an image chunk) in one go. This speeds up training and encourages more cohesive outputs.

Training Data Mix: One key to BAGEL’s smarts is its interleaved multimodal pretraining . Rather than train separately on text and image corpora, the team mixed them. A single training sequence could jump from a few lines of text to an image to a video frame, then back to text. This “multimodal diet” teaches the model how modalities relate. For example, it might learn that the sentence “A dog is playing fetch” often appears with images of dogs and tennis balls. According to the authors, this approach unlocks surprising abilities (emergent properties) that purely text- or image-trained models don’t learn.

Open-Source Release: Importantly, BAGEL is fully open-source under Apache 2.0. ByteDance-Seed has shared the weights on Hugging Face and GitHub, along with training details and data protocols. This contrasts with models like GPT-4o (OpenAI) or Gemini, which are proprietary. The open license means anyone – startups, researchers, even hobbyists – can download BAGEL, fine-tune it, or plug it into applications without paying fees.

BAGEL vs. GPT-4o (A Quick Comparison)

It’s natural to compare BAGEL with other high-profile multimodal models. One obvious rival is GPT-4o (sometimes called GPT-4 Omni), which is OpenAI’s multimodal flagship. The key difference is openness: GPT-4o is closed-source and only accessible via API, whereas BAGEL is open-source (Apache 2.0) and self-hostable. Functionally, both aim to “understand and generate text, images, and video,” but BAGEL even extends to tasks like 3D reasoning and outpainting that aren’t (yet) GPT-4o features. In short, BAGEL offers a comparable multimodal skill set without any paywall , making it a “GPT-4o alternative” for the community. (Of course, GPT-4o may be trained on vastly more data or have other proprietary optimizations, so actual performance might differ.) The point is: if you need a multimodal AI that you can fully control and modify, BAGEL is now a viable option.

Getting Started with BAGEL

For developers excited to try BAGEL, ByteDance has made it fairly easy. The model weights and code are on GitHub/Hugging Face, and there’s even an online demo. For example, ByteDance launched a HuggingFace Space with a simple UI where you can upload an image or enter text and see BAGEL’s analysis and generation in action. This means you can experiment with BAGEL without powerful hardware on your end. (If you do have a decent GPU, you can clone their GitHub repo and run it locally; they provide a quick-start guide and pretrained checkpoint in under 50k lines.)

Because BAGEL is open, you can also fine-tune or extend it for your own projects. For instance, you might want to specialize BAGEL for medical images or adapt it to specific video tasks. In such cases, it’s often useful to prepare your custom data as text prompts. Tools like Repo2Txt can help here. If you have a code repository or dataset (say, a GitHub project) that your model should learn from, you can convert a GitHub repo to text using our GitHub Repo Converter, or turn a local folder into a single text file with our Local Repo Converter. (This repo-to-text process is exactly what LLM developers use to streamline code-based training data.) Similarly, if you need to pull data from websites or docs, our Crawl4AI (Web2Txt) tool can scrape pages and deliver text for LLM prompts. In short, any code or content you want BAGEL to reason about can be easily prepared.

Integration & Services: Incorporating BAGEL into real-world applications may require engineering effort (especially for large-scale inference). That’s where AI/ML consultancies and dev shops come in. For example, companies like Its IT Group (itsitgroup.com) specialize in end-to-end AI/ML solutions, web and app development, and DevOps. A team like that could help deploy BAGEL on servers, integrate it with apps (chatbots, image tools, etc.), and optimize performance. If you’re exploring BAGEL for a product, partnering with a services firm is one way to accelerate development.

Why BAGEL Matters

BAGEL represents a big step for open-source AI. It shows that an open model can match or exceed many tasks of closed models – and do so transparently. For developers and researchers, that’s empowering. You can prototype multimodal apps faster (no waiting for API access), trust what the model is really doing, and even customize it. For instance, some startups might build creative design tools on top of BAGEL, letting users seamlessly edit images with natural language. Others may use it for data analysis (e.g., summarizing charts or video logs).

Moreover, BAGEL’s open release sets a precedent. The AI community can now study its “brain”: researchers can analyze why BAGEL handles 3D reasoning or where it fails. They can gather “bad cases” and improve the model collaboratively. This contrasts with black-box models: with GPT-4o you don’t see the internals. With BAGEL on GitHub, it’s the whole package.

Finally, BAGEL’s arrival underscores the hot pace of AI development in 2025. Just weeks before, we saw things like GPT-4.5 on the horizon, and new models for 3D asset creation (e.g. the Hunyuan 3D model). It’s an exciting time: our 2025 LLM release timeline shows dozens of big releases in just a few months. BAGEL fits right into this picture – a cutting-edge, open-source multimodal model that the community can build upon.

In summary, if you’re a developer or AI enthusiast, BAGEL is worth exploring today . It can handle image and video tasks you’d normally give to separate tools, and you can run it yourself or via demos. And remember, preparing your data is part of the game: whether it’s converting a GitHub project to text for fine-tuning or scraping web data, tools like [repo2txt] help bridge the gap between raw content and LLM-ready input. With models like BAGEL and the ecosystem around them, multimodal AI is becoming more accessible than ever.