--- license: mit base_model: empero-ai/Qwythos-9B-v2 language: - pt - en pipeline_tag: text-generation library_name: llama.cpp tags: - MoE - Qwen - CobrIX-1.0-Coder-Flash - custom-architecture - pytorch - causal-lm - GGUF - quantization --- # CobrIX-1.0-Coder-Flash-MoE — GGUF **CobrIX-1.0-Coder-Flash-MoE** is a custom **Mixture-of-Experts (MoE)** coding model distributed in ready-to-use **GGUF** format. The model is built from the dense Qwen 3.5 base model **`empero-ai/Qwythos-9B-v2`** and incorporates **5 specialized dense fine-tuned experts**. The GGUF files provided in this repository are already quantized and can be used directly with compatible inference applications such as **llama.cpp, LM Studio, and other GGUF-compatible runtimes**. --- ## 🚀 Download Choose the quantization according to your available RAM/VRAM and desired quality. | File | Quantization | Approx. BPW | Recommended For | | --------------------------------------------- | -----------: | ----------: | -------------------- | | `CobrIX-1.0-Coder-Flash-33B-A13B-Q3_K_M.gguf` | Q3_K_M | ~3.5 bpw | Low-VRAM systems | | `CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf` | Q4_K_M | ~4.8 bpw | ⭐ Best balance | | `CobrIX-1.0-Coder-Flash-33B-A13B-Q5_K_M.gguf` | Q5_K_M | ~5.5 bpw | Higher quality | | `CobrIX-1.0-Coder-Flash-33B-A13B-Q8_0.gguf` | Q8_0 | ~8.5 bpw | Maximum GGUF quality | ### ⭐ Recommended For most users, **Q4_K_M** is the recommended version: ```text CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf ``` It provides a strong balance between model quality, memory usage, and inference performance. --- # 🧠 Model Overview CobrIX-1.0-Coder-Flash-MoE is designed as an efficient **AI coding copilot**, optimized for interactive development and human-supervised workflows. ### Model characteristics * **Architecture:** Mixture-of-Experts (MoE) * **Total parameters:** ~33B * **Active parameters:** ~13B per token * **Experts:** 5 * **Experts activated per token:** 2 * **Shared expert:** 1 * **Context window:** up to 1,048,576 tokens * **Format:** GGUF * **Primary use:** Coding and technical reasoning * **Languages:** Portuguese and English --- # ⚡ Capabilities ## 🧩 Advanced Reasoning The Flash model is designed for complex reasoning, problem decomposition, debugging, and technical problem solving. It is particularly suited to situations where the model works alongside a human developer as an interactive AI copilot. ## 💻 Expert Programming Strong focus on software development, including: * Python * JavaScript * TypeScript * React * Next.js * Node.js * Backend development * Frontend development * APIs * Database systems * Debugging * Refactoring * Software architecture * Automation * DevOps ## 🔐 Cybersecurity The model also provides capabilities for: * Security-oriented code analysis * Vulnerability analysis * Secure programming * Security scripting * Understanding common application-security concepts The primary focus of the model remains **software engineering and programming**. --- # 🏗️ Architecture The model uses a custom MoE architecture based on the Qwen 3.5 model family. ### Model class ```text Qwen35MoEForCausalLM ``` ### Model type ```text qwen35_moe ``` ### Context window ```text 1,048,576 tokens ``` Configuration: ```text max_position_embeddings = 1048576 ``` ### Expert routing The model contains: ```text 5 total experts 2 experts activated per token ``` This uses **Top-2 routing**, where the router selects two experts from the five available experts for each token. ### Shared expert The architecture also contains one shared expert that remains active for every token. ```text num_shared_experts = 1 ``` The shared expert is based on the base model's MLP and uses a sigmoid gating mechanism. --- # 📦 GGUF Quantizations ## Q3_K_M ```text CobrIX-1.0-Coder-Flash-33B-A13B-Q3_K_M.gguf ``` The smallest provided quantization. Recommended when memory usage is the primary concern. **Advantages:** * Lowest memory requirements * Good performance on lower-end hardware * Suitable for systems with limited VRAM/RAM **Trade-off:** * Lower quality than Q4/Q5/Q8 --- ## Q4_K_M ⭐ Recommended ```text CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf ``` The recommended general-purpose quantization. **Advantages:** * Excellent quality/size balance * Good inference speed * Suitable for a wide range of GPUs and CPUs * Recommended for everyday coding --- ## Q5_K_M ```text CobrIX-1.0-Coder-Flash-33B-A13B-Q5_K_M.gguf ``` Provides higher precision than Q4_K_M. Recommended when additional memory is available and maintaining higher model quality is more important than minimizing model size. --- ## Q8_0 ```text CobrIX-1.0-Coder-Flash-33B-A13B-Q8_0.gguf ``` The highest-precision GGUF quantization provided in this repository. Recommended for systems with sufficient memory where maximum GGUF quality is preferred. --- # 🦙 Using with llama.cpp The GGUF files can be loaded directly using **llama.cpp**. Example using `llama-cli`: ```bash llama-cli \ -m CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf \ -p "Write a Python script to set up an async web server using FastAPI." \ -n 512 \ -ngl 99 \ -c 4096 ``` ### Important The `-c 4096` value in this example is only an example context size. The model supports a native context window of up to: ```text 1,048,576 tokens ``` However, using the maximum context requires significantly more memory and depends on the inference backend and hardware. --- # 🐍 Python — llama-cpp-python Install the Python binding: ```bash pip install llama-cpp-python ``` Then: ```python from llama_cpp import Llama llm = Llama( model_path="./CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf", n_gpu_layers=-1, n_ctx=4096, verbose=True ) output = llm( "Hello, how can I optimize a React re-render loop?", max_tokens=128, stop=["<|endoftext|>"], echo=False ) print(output["choices"][0]["text"]) ``` --- # 🖥️ LM Studio The GGUF files can also be loaded into **LM Studio**. ### Steps 1. Download the desired `.gguf` file from this repository. 2. Open LM Studio. 3. Import or place the GGUF file in your local model directory. 4. Select the model. 5. Configure the desired GPU offloading and context size. 6. Start the model. For most users, start with: ```text Q4_K_M ``` and increase the quantization if additional memory is available. --- # 🦙 Ollama The GGUF weights can also be used with compatible Ollama workflows. A basic `Modelfile` can be created as follows: ```text FROM ./CobrIX-1.0-Coder-Flash-33B-A13B-Q4_K_M.gguf ``` Then create the model: ```bash ollama create cobrix-1.0-coder-flash -f Modelfile ``` Run it: ```bash ollama run cobrix-1.0-coder-flash ``` --- # 📊 Quantization Selection A simple recommendation: ```text Limited VRAM/RAM │ ▼ Q3_K_M │ ▼ Q4_K_M ⭐ │ ▼ Q5_K_M │ ▼ Q8_0 │ ▼ Higher quality / higher memory usage ``` For general-purpose usage: > **Q4_K_M is the recommended starting point.** --- # 🎯 Intended Use CobrIX-1.0-Coder-Flash-MoE is intended primarily for: * Software development * AI-assisted programming * Code generation * Debugging * Refactoring * Technical reasoning * Application development * Web development * Backend development * Frontend development * Automation * Developer assistance The model is designed to work best as a **human-supervised coding copilot**, where the developer reviews and validates generated code. --- # ⚠️ Disclaimer Generated code should always be reviewed and tested before being used in production environments. The model may produce incorrect, outdated, insecure, or hallucinated information. Do not blindly execute generated commands or deploy generated code without appropriate validation and security review. --- # 📜 License This repository is released under the **MIT License**. Please also review the licensing and terms of the underlying base model and any models or datasets used during the creation of the experts. --- ## CobrIX Coder and CobrIX Code Models — Early Access **CobrIX Coder** is the AI ​​ecosystem that runs and provides access to models such as **CobrIX-1.0-Coder-Flash-MoE** and **CobrIX-1.0-Coder-Full-MoE**. 🌐 **Check out and test the AI:** https://cobrix.vercel.app/coder Visit the site to explore the AI ​​and follow the development of **CobrIX Code**. ### Interest List If you want to know when the AI ​​ecosystem and **CobrIX Code** will be available to the public, visit: **https://cobrix.vercel.app/coder** On the site, you will find a field to enter your email and join the **early access waitlist**. Registered users will receive an **email when the CobrIX AI ecosystem and CobrIX Code become available**. We currently have the funds to host CobrIX-1.0-Coder-Flash-33B-A13B on an RTX 6000 Ada Generation GPU (48GB VRAM), supporting between 50 and 100 concurrent users. Please help us by donating (links below); help us create AI that is usable for everyone—open-source AI forever—so we can evolve together. > 🚀 Join the waitlist and get notified when CobrIX Code is released. # ❤️ Support CobrIX Developing and maintaining custom AI architectures such as **CobrIX-1.0-Coder-Flash-MoE** requires computing resources, model development, experimentation, infrastructure, and ongoing maintenance. If this project is useful to you or your organization, you can support the development of open AI infrastructure and future model releases through the wallets below. ### Bitcoin (BTC) ```text bc1q8mu8fjak4y84qj4dlk8pu4d3zhknm92zra4r4m ``` ### Ethereum (ETH / ERC-20) ```text 0x8D9187dEa0a77390ef668361cd5b236DE54af2BB ``` ### Solana (SOL) ```text GQR2jZnWuWP1c3dbuz4mC7ZnyveacBKy63q8qf9nj8bp ``` All donations help support: * AI infrastructure * Compute resources * Open-source model development * Model experimentation * Future CobrIX releases * Community support Thank you to all developers, researchers, enthusiasts, and contributors supporting the development of open and accessible artificial intelligence. --- # 🌎 CobrIX **CobrIX-1.0-Coder-Flash-MoE** > An efficient open AI coding copilot powered by a custom Mixture-of-Experts architecture.