Projects

Overview: This is a DeepSeek-V4-Flash-0731 MXFP4 experiment branch specifically designed for NVIDIA RTX 4090 (SM89) with 8×RTX 4090 24 GB. The repository contains only the inference framework source code, not including model weights, JIT cache, benchmark requests, access credentials, or server-specific configurations. Key features: - MXFP4 routed experts - Target-only mode with decode full CUDA Graph - Kernel optimizations for SM89 architecture - Performance benchmarks for 8×RTX 4090 24 GB with TP=8 Tech stack: Python Why it matters: This branch is designed for a single machine with 8 NVIDIA RTX 4090 GPUs, targeting the Ada SM89 architecture for inference. It includes optimizations for the SM89 GPU, including kernel routing for MXFP4 Marlin MoE, sparse decode, and fallbacks for certain components. The performance benchmarks show an efficiency of 87.5% for the Scheduler output, 85.5% for the decode window, and 72.7% for the end-to-end aggregation, indicating significant improvements over the current production version. GitHub stars: 8

### Overview: This repository contains 8 tested FLUX 2 Pro prompts, each with an example render, aspect ratio, and the cost per image. The prompts are designed to help developers create high-quality image recipes efficiently. ### Key Features: - **Model ID:** `flux-2-pro` - **Prompt Structure:** - **Subject:** Concrete objects or people with specific materials. - **Lighting:** Direction and quality of light. - **Framing:** Ratio and lens language. - **Setting:** Surface, backdrop, or environment. - **Finish:** Render or film character. - **Effective Price:** $0.024 per image, estimated cost for the set: $0.192. ### Tech Stack: Python ### Why It Matters: - **Access to Multiple Models:** The same base URL and authorization header allow access to FLUX 2 Pro and over 300 image, video, and language models. - **Cost-Effective Pricing:** No subscription or upfront plan, pay-as-you-go pricing with immediate recharges. - **Real-Time Cost Reporting:** Each completed task reports the cost, and failed tasks are refunded. - **Asynchronous Nature:** Tasks can be sent, and the `task_id` is returned, allowing for batch processing and retries through background tasks. ### What Separates a Usable Recipe from a Wasted Call: - **Material vs. Adjective:** "Frosted glass" vs. "beautiful" or "high quality". - **Light Direction:** "Single softbox

Overview: Seedance 2.0 Mini Prompts is a collection of 4 tested image prompt recipes designed for developers, showcasing effective and efficient use of the Seedance 2.0 Mini model. Each recipe includes a concrete subject, light direction and quality, framing ratio, setting surface, and finish, ensuring clear and compelling outputs. Key features: - 4 tested image prompt recipes - Example renders, durations, and call costs on the Seedance 2.0 Mini model - Effective price of $0.066 per second - Supports OpenAI-compatible endpoint - No need for monthly subscription - Provides detailed prompt structure guide Tech stack: Python Why it matters: Seedance 2.0 Mini Prompts offers developers a concise and effective way to use the Seedance 2.0 Mini model, ensuring clear and compelling image outputs at an affordable cost. Repository: y-agn7/seedance-2-0-mini-prompts-ko Primary language: Python GitHub stars: 8 GitHub description: Seedance 2.0 Mini Prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-2.0-mini`) and the cost per image.

Overview: Nano Banana 2 API provides a cost-effective and OpenAI-compatible endpoint for image generation, featuring a single endpoint that supports multiple resolutions and pricing models. Key features: - Single endpoint for image generation - Supports multiple resolutions (1K, 2K, 4K) - Pricing models: $0.015 per image, $0.025 for 4K - Assynchronous execution with task tracking Tech stack: Python Why it matters: Nano Banana 2 API simplifies API management by providing a unified endpoint for image generation, reducing complexity and improving efficiency. Repository: ydkd8kdkl8n/nano-banana-2-api-pt Primary language: Python GitHub stars: 8 GitHub description: Nano Banana 2 API (`gemini-3.1-flash-image-preview`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: Maestro AMD is a 100% local AI video, image, and music studio that runs on AMD GPUs, providing a seamless creative experience without cloud or subscription requirements. Maestro, the underlying AI tool, is built for NVIDIA but this wrapper adds AMD support, making it accessible to AMD GPU owners. ### Key Features: - **Director Mode**: Generates video, images, and audio with prompts, handling everything locally. - **Studio Mode**: Offers full manual control over the underlying models, including video, image, and audio. - **Edit Mode**: Provides tools for retakes, edits, and outpainting, allowing for post-production adjustments. - **Performance Auto-Tune**: Automatically adjusts quality and settings based on GPU and RAM detection. - **Built-in Local LLM**: The planning brain runs on PyTorch built for ROCm, ensuring compatibility with AMD hardware. - **CivitAI LoRA Browser**: Allows users to search, install, and manage LoRAs within the app. - **Workspaces**: Keeps projects and output folders separate, ensuring isolation. - **Dark/Light Themes**: Available options for adjusting the app's appearance. - **NSFW Mode**: Optional feature for users who prefer a more mature environment. ### Tech Stack: Maestro AMD uses PyTorch built for ROCm to run the same models and features as Maestro, ensuring compatibility with AMD GPUs. ### Why It Matters: Maestro AMD offers a complete creative solution for AMD GPU owners, providing a local, cloud-free experience for generating video, images, and audio

Overview: This repository contains Nano Banana 2 Lite prompts with example renders, aspect ratios, model IDs (`gemini-3.1-flash-lite-image-ext`), and cost per image. The prompts are tested and can be used for various image and video tasks. Key features: - 8 tested recipes with example renders, aspect ratios, model IDs, and cost per image. - Nano Banana 2 Lite prompts are machine-readable in `data/prompts.json` and table in `PROMPTS.md`. - The effective price used is $0.0125 per image, with an estimated cost of $0.1 for the entire set. Tech stack: Python Why it matters: This repository provides a practical guide for developers looking to use Nano Banana 2 Lite prompts effectively, ensuring efficient and cost-effective image and video tasks.

### Overview: Nano Banana 2 Lite Prompts is a collection of 8 tested image prompt recipes, showcasing examples, aspect ratios, and the cost per image using the OpenAI-compatible endpoint `gemini-3.1-flash-lite-image-ext`. This resource is designed to help developers and artists create visually appealing and cost-effective images. ### Key Features: - **8 Recipes**: Each recipe includes a concrete object or person, material, light direction, framing ratio, and setting. - **Example Renders**: Detailed render examples are provided for each recipe. - **Cost Calculation**: The cost per image is given, with an estimated total cost for the set. - **Model Compatibility**: Accessible through a base URL and `Authorization` header, allowing for easy integration. - **Subsription-Free**: No need for a monthly subscription or pre-paid plans. - **Pricing Transparency**: The cost is reported in the response, with returns for failed tasks. - **Non-Blocking Design**: Uses non-blocking design principles, allowing for efficient batch processing. ### Tech Stack: Python ### Why It Matters: Nano Banana 2 Lite Prompts offers a straightforward and effective way to create visually appealing images without the need for complex setup or expensive subscriptions. The recipes are designed to help developers and artists create high-quality images efficiently and cost-effectively.

Overview: DaoQL-Edu is a simplified, educational implementation of the DaoQL multimodal data engine, designed for database systems courses and self-learners. Key features: Multi-Engine Unified, DSL Query Language, Cross-Engine Nested Queries, SIMD Acceleration, HNSW Vector Index, Transaction & WAL. Tech stack: Rust 1.78+. Why it matters: DaoQL-Edu provides a clear and educational view of the DaoQL engine's design and implementation, ideal for database systems courses and self-learners. Repository: zhanbolee/DaoQL-Edu GitHub stars: 8 GitHub description: A simplified, educational multimodal data engine (Graph + Column + Vector) in pure Rust — ideal for database systems courses and self-learners.

### Overview: The 1bit engine is a model-agnostic local language model inference solution designed for the AMD Ryzen AI MAX+ 395 (Strix Halo) on Linux. It leverages Radeon 8060S iGPU via HRX with HIP kernels and XDNA 2 NPU via our NPU runtime, running on a backend of Lemonade Server with an OpenAI-compatible API. This setup allows for efficient inference of any GGUF on the GPU and NPU. ### Key Features: - **Model Agnostic**: Supports any GGUF model. - **GPU and NPU Backend**: Inference runs on the GPU and NPU. - **HIP and NPU Runtime**: Utilizes HIP kernels and our NPU runtime for optimal performance. - **Lemonade Server Integration**: Integrated with Lemonade Server for seamless integration and compatibility with OpenAI APIs. ### Tech Stack: C++ ### Why it Matters: The 1bit engine simplifies the process of deploying large language models locally on AMD hardware, making it accessible and efficient for developers and researchers. Its model-agnostic nature and GPU/NPU backend ensure flexibility and performance, making it a valuable tool for various applications and research projects.

Overview: Block is a prototype framework designed for balancing load in large-language-model (LLM) serving, focusing on predictive, performance-aware scheduling. Key features: - Predictive, performance-aware scheduling - Randomized power-of-two-choices (Po2) predictive dispatch - Light-weight length-estimation model for handling prompts unseen in training Tech stack: Python Why it matters: Block addresses the challenge of efficient and load-aware scheduling in LLM inference, demonstrating improvements in probing cost and herding under bursty arrivals, while matching baseline serving capacity. Repository: AKafakA/Block Primary language: Python GitHub stars: 7 GitHub description: Prototype for Block, a new LLM-multi instance scheduling Framework README excerpt: ...

### Overview: Isvik is a high-performance, ultra-lightweight, single-binary local AI inference runtime platform built 100% in Nim. Leveraging the low-level library Resonance to interface directly with the OpenVINO C API, Isvik is engineered specifically for Intel AI PCs (Intel Core / Core Ultra) to provide unified, deeply optimized local AI inference across CPU, iGPU (Arc, Iris Xe), and NPU (Intel AI Boost). ### Key features: - **100% Native Pure Compiled Binary**: No bulky Python virtual environments, no Node.js/Electron overhead. Sub-100ms cold starts with minimal memory and system footprint. - **Intel Hardware Auto-Scheduling**: Intelligent `AUTO` device pipeline (`NPU → Intel GPU → CPU`), precise vendor discrimination (`0x8086` Intel only, excluding discrete NVIDIA GPUs from Intel targets), with automated CPU fallback on iGPU single memory allocation limits (4 GiB). - **Interactive Terminal Workbench & REPL**: Modern interactive shell with numbered model selector (`[1] [2]`), one-click model swapping, bilingual interface (`zh-CN` / `en-US`), session memory, and streaming generation. - **Unified Multi-Protocol Local API Server**: Built-in HTTP server with native compatibility for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages protocols. Drop-in ready for TRAE IDE, Continue.dev, Open WebUI, Cherry Studio, and custom clients. - **Industrial-Grade Safety**: Safe opaque integer handles (zero

Overview: This repository contains Team Envisage's complete submission to the DL Sprint 4.0 competition, organized as part of BUET CSE Fest 2026. The competition consisted of two challenging Bengali speech processing tracks: Bengali Long-Form ASR and Bengali Speaker Diarization. The fine-tuned Bengali diarization model is publicly available on Hugging Face, and a live demo is available for interactive use. Key features: - Published model: Our Bengali diarization model is available on Hugging Face. - Live demo: Try our Bengali Speaker Diarization system interactively. - IEEE conference paper: The full research paper is included in the repository. Tech stack: PyTorch, pyannote.audio, and Hugging Face. Why it matters: This project highlights the underrepresentation of Bengali in long-form speech technology and demonstrates the potential of fine-tuned models in Bengali speech processing. The live demo offers an interactive way to experience the system, making it accessible to a broader audience. GitHub stars: 7 Repository: AdilShamim8/BUET-CSE-Fest-2026

Overview: ChatPDF is a document interaction tool that allows users to interact with PDF documents using AI and LangChain, providing a user-friendly interface for querying and extracting information. Key features: - Upload multiple PDF files - Ask specific questions about the content - Receive detailed responses in real-time Tech stack: Python, LangChain, Streamlit, FAISS, MIT License Why it matters: ChatPDF simplifies the process of interacting with PDF documents, making it easier for users to extract information and answers directly from their documents.

### Overview: This project is a one-command recipe for running GLM-5.3-Flash on two 96 GB RTX PRO 6000 Blackwell GPUs, utilizing TensorFold v0.6.6, Mia's AI Lab's EXL3 checkpoint, and Aevonix Research's optimizations. It supports text and vision tasks and can handle up to 4 concurrent requests with 524,288 tokens per request. ### Key Features: - **Checkpoint:** `Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold` - **Drafter:** `incoai/GLM-5.3-Flash-DFlash2` - **API Model ID:** `glm-5.3-flash` - **Context Tokens:** 524,288 tokens per request - **Concurrency:** 4 requests by default, with a 32-row verification window - **One Command:** `./start.sh`; restart with `./start.sh restart`, stop with `./stop.sh` ### Tech Stack: Python ### Why It Matters: This recipe provides a streamlined, efficient way to run GLM-5.3-Flash on a Linux host with two RTX PRO 6000 GPUs, leveraging TensorFold for enhanced performance and Mia's AI Lab's checkpoint for improved results. The recipe is designed to be easy to use with a single command, making it ideal for developers and researchers looking to integrate GLM-5.3-Flash

### Overview: This project is a production-grade genpark-beam-search-autoregressive-decoder-skill skill designed for AI agents. It provides a pure standard library Python implementation with zero pip dependencies, focusing on fundamental computational linguistics and NLP algorithms such as Earley, CKY, BPE, Beam Search, and BM25. The project supports native Model Context Protocol (MCP) server for AI agent text intelligence. ### Key features: - Pure standard library Python implementation - Fundamental computational linguistics and NLP algorithms - Native Model Context Protocol (MCP) server support - Supports AI agent text intelligence ### Tech stack: Python ### Why it matters: This skill is essential for AI agents looking to implement advanced sequence decoding techniques, ensuring high performance and efficiency in language processing tasks.

Overview: `genpark-dynamic-prompt-prefix-cache-skill` optimizes edge AI inference by eliminating redundant prefill computation and reducing total token time to fill (TTFT) through Radix trie-based prompt prefix caching. Key features: - Zero external dependencies - Implements formal INT8 symmetric quantization scales - Non-contiguous PagedAttention virtual block mapping - Speculative decoding rejection sampling - Radix trie prompt prefix caching - High-precision TTFT/TPOT latency telemetry - Native Anthropic MCP Protocol Tech stack: Python 3.9+ Why it matters: This skill provides zero-dependency, mathematically sound edge AI acceleration and inference optimization primitives engineered strictly using Python 3.9+ standard library, ensuring production-grade performance and reliability.

Overview: This project is a zero-dependency, mathematically sound edge AI acceleration and inference optimization primitive managed exclusively via Python 3.9+ standard library. It includes zero external dependencies, enterprise edge invariants, and native Anthropic MCP protocol compliance. Key features: - Zero External Dependencies: Operates exclusively via pure Python. - Enterprise Edge Invariants: Implements formal INT8 symmetric quantization scales, non-contiguous PagedAttention virtual block mapping, speculative decoding rejection sampling, Radix trie prompt prefix caching, and high-precision TTFT/TPOT latency telemetry. - Native Anthropic MCP Protocol: Compliant with standard JSON-RPC 2.0 stdio MCP specifications for Claude Desktop, Cursor, and Windsurf. Tech stack: Python 3.9+ Why it matters: This skill is designed for production-grade edge AI acceleration and inference optimization, ensuring efficient and accurate inference with minimal external dependencies.

### Overview: The Grok Imagine Image 2.0 API and Grok Imagine Image 2 API Python SDK provide a powerful toolset for generating and editing images using MuAPI. This comprehensive solution allows developers to easily integrate xAI image generation and editing functionalities into their projects. ### Key Features: - **Python SDK**: A Python library that simplifies the integration of the Grok Imagine Image 2.0 API into Python applications. - **MCP Server**: A server-side component that handles requests and manages asynchronous job polling, ensuring efficient and scalable image processing. - **MuAPI Integration**: Seamless integration with MuAPI, a unified API for image, video, and audio generation, making it easy to access a wide range of AI models. ### Tech Stack: Python ### Why It Matters: This solution is crucial for developers looking to integrate advanced image generation and editing capabilities into their projects. The Python SDK makes it straightforward to integrate the API into Python applications, while the MCP server ensures efficient and scalable image processing. MuAPI's unified API makes it easy to access a wide range of AI models, making this solution highly versatile and powerful. ### Repository: Anil-matcha/Grok-Imagine-Image-2-API GitHub Stars: 7 GitHub Description: Grok Imagine Image 2.0 API and Grok Imagine Image 2 API Python SDK and MCP server for xAI image generation and editing through MuAPI. License: MIT Primary Language: Python GitHub URL: https://github.com/Anil-matcha/Grok-Imagine

Overview: This project provides custom ComfyUI nodes for generating high-quality Midjourney images using the muapi.ai API, including V7, V8, and Niji models. Key features: - Midjourney V7 - Midjourney V8 - Midjourney Niji Tech stack: Python Why it matters: This allows developers to integrate advanced image generation capabilities directly into ComfyUI, enhancing their workflow with high-quality Midjourney images. Repository: Anil-matcha/midjourney-comfyui GitHub stars: 7 GitHub description: ComfyUI custom nodes for Midjourney V7, V8, and Niji via muapi.ai — 4 images per run, reference image support, one x-api-key header.

Overview: A versatile Telegram bot offering daily meditation prompts, AI chat, media downloads, and channel membership verification in both English and Persian. Key features: - Daily Meditation Prompts - AI Chatbot Interaction - Media Downloads - Channel Membership Verification Tech stack: Python Why it matters: This bot provides a comprehensive service for users seeking mindfulness, AI assistance, media access, and community engagement through Telegram. Repository: ArmanShirzad/telegram-bot-to-checkmembers-pub GitHub stars: 7