Projects

**Overview:** Unified static-image generation routing for DeepSeek Harness, featuring a single tool that supports multiple engines with fallback health checks and durable results. **Key features:** - Health-aware fallback - Durable results - Counted usage **Tech stack:** TypeScript **Why it matters:** This plugin simplifies the static-image generation process for DeepSeek Harness, ensuring reliable and efficient image creation. **Repository:** PerryLink/dsh-draw **Primary language:** TypeScript **GitHub stars:** 11 **GitHub description:** Unified static-image generation router for DeepSeek Harness: one image_generate tool with standard parameters, config-driven engine routing (OpenAI Images, Zhipu CogView, any compatible endpoint) with health-aware fallback, durable workspace attachments and per-session quota accounting. Maintenance status: FROZEN since 2026-10-05.

Overview: This repository contains 150 production-ready prompts for OpenAI's GPT Image 2 model, covering categories such as cinematic portraits, action & sports, nature & landscapes, fantasy & illustration, product & commercial, UI/UX design, game assets, character design, poster & graphic design, and text rendering. Key features: - 150+ production-ready prompts - 15 categories organized by category - 150+ examples of prompts and results Tech stack: HTML Why it matters: This repository provides a curated collection of prompts for designers and developers to use in their projects, offering a wide range of styles and applications for various design and creative projects.

Overview: A Telegram bot that leverages Large Language Models (LLMs) and Playground AI to generate AI images. Key features: - Generates images based on LLM prompts - Supports Playground AI for enhanced image generation - Can be used to create response images for various prompts Tech stack: JavaScript Why it matters: This bot streamlines the process of generating AI images, making it easier for developers to create engaging and visually appealing content without needing deep expertise in AI or image generation tools.

### Overview: Imference Desktop is a free, one-click desktop app that allows you to generate AI images and videos on your own GPU or in the cloud. It supports Windows and macOS, and includes models like SDXL, SD 1.5, Z-Image, FLUX, Chroma, Qwen-Image, Anima, and Krea 2. ### Key Features: - **One-click Setup**: The app installs an isolated inference engine, detects your GPU, and starts and stops the engine on demand. - **Local Generation**: Run eight model families on your own GPU for free, including SDXL, SD 1.5, Z-Image, FLUX, Chroma, Qwen-Image, Anima, and Krea 2. - **Preconfigured Models**: Pick a model from a curated catalog and hit Generate, where the weights automatically download and the models are pre-tuned. - **LoRAs**: Stack up to four styles, characters, or detail boosters on SDXL. ### Tech Stack: Go ### Why It Matters: - **Ease of Use**: The app is one-click setup, installs an isolated inference engine, and detects your GPU. - **Cost-Effective**: Run eight model families on your own GPU for free, and your prompts and images never leave your machine. - **Preconfigured Models**: Pick a model from a curated catalog and hit Generate, where the weights automatically download and the models are pre-tuned. - **LoRAs**: Stack up to four styles, characters, or detail boosters on SD

### Overview: Gridbook is a format specification and out-of-tree vLLM plugin designed to serve 2-6 bit-per-weight LLM quantization weights on native Blackwell tensor cores. This allows for native performance without the need for forked runtimes or custom serving stacks. ### Key Features: - **Format Specification**: Gridbook serves weights as product-vector-quantization codebook indices, ensuring that the decoded tiles are exactly what the hardware expects. - **Out-of-Tree vLLM Plugin**: This plugin is designed to work seamlessly with existing vLLM models, allowing for the efficient serving of quantized weights. - **Compatibility**: Gridbook supports NVIDIA Blackwell `sm_120` / `sm_121` GPUs and is compatible with various versions of PyTorch and vLLM. ### Tech Stack: - **Python**: 3.10+ - **CUDA Toolchain**: `nvcc` on `PATH` - **PyTorch**: 2.11.0+cu130 - **vLLM**: 0.23.1rc1.dev764+g54b16d8a9 ### Why It Matters: Gridbook provides a straightforward solution for quantized LLM weights, ensuring that the weights are efficiently served on native hardware without the need for complex runtime modifications. This makes it ideal for developers and researchers looking to optimize their models for performance on NVIDIA hardware.

### Overview: Muapi MCP Server is a project designed to integrate MuAPI with MCP-compatible AI assistants, allowing clients like Claude, Cursor, and others to access a wide range of generative AI models directly from MuAPI. This project simplifies the process of setting up MuAPI for use with MCP clients, making it easier to integrate these powerful AI assistants into various applications. ### Key Features: - **MuAPI Integration**: Connects any MCP-compatible AI assistant to MuAPI, enabling access to 100+ generative AI models. - **Server Setup**: Provides a self-hosted MCP server instance for those who prefer to run their own infrastructure. - **Client Configuration**: Serves as a template for configuring MCP clients to point to the MuAPI server. ### Tech Stack: - Python ### Why It Matters: Muapi MCP Server simplifies the integration of MuAPI with MCP-compatible AI assistants, making it easier to incorporate these powerful AI tools into various applications. The project is designed to be user-friendly and straightforward, making it accessible to developers and enthusiasts alike. ### GitHub Stars: 12 ### Repository: - **SamurAIGPT/muapi-mcp-server** ### Primary Language: - Python ### Related Projects: - **seedance-2-mcp**: MCP server for Seedance 2 video generation with model-specific schemas. - **seedance-2.5-mcp**: MCP server for Seedance 2.5 Preview with 720p/480p route selection. - **muapi-skills**: Agent skills for Claude

### Overview: This project focuses on fine-tuning the Qwen2.5-7B-Instruct model for a hotel domain, aiming to create a practical, reproducible, and stable hotel service assistant. It includes SFT + LoRA, SFT + QLoRA, and DPO techniques, providing a comprehensive solution for AI engineers. ### Key Features: - **SFT-LoRA**: Low-risk baseline for stable training. - **SFT-QLoRA**: Low-memory training for 7B model. - **DPO**: Enhances responses through human preference pairs. ### Tech Stack: Python ### Why It Matters: This project offers a practical, reproducible, and stable solution for hotel domain conversational assistants, making it suitable for AI engineers looking to improve their models. Repository: SuleynanAuir/Distributed-Fine-Tuning-of-Qwen2.5-7B-as-HotelService-Master Primary Language: Python GitHub Stars: 12

Overview: This project introduces a method for detecting satellite and instrument entities in unstructured text, leveraging pre-trained language models and distant supervision. The models are available for download and online testing, aiding in the accurate extraction of Earth observation resources. Key features: - Pre-trained language model for satellite and instrument entity detection - Distant supervision for improved accuracy - Available models: satellite-instrument-roberta-NER, satellite-instrument-bert-NER Tech stack: Python, Hugging Face Transformers Why it matters: This approach helps in linking and reusing Earth observation resources by accurately extracting satellite and instrument entities from unstructured texts, addressing the challenge of linking and reusing Earth observation resources. Repository: THU-ESIS/Satellite-Instrument-NER Primary language: English GitHub stars: 12 GitHub description: Earth observations, especially satellite data, have produced a wealth of methods and results in meeting global challenges, often presented in unstructured texts such as papers or reports. Accurate extraction of satellite and instrument entities from these unstructured texts can help to link and reuse Earth observation resources.

### Overview: Today-Hbw/rag-platform is an open-source RAG platform designed for pluggable document ingestion, multimodal indexing, hybrid search, tenant-isolated APIs, RBAC, citations, and MCP-based AI agent integration. It separates ingestion, indexing, retrieval, lifecycle management, and agent integration into independently deployable services, supporting dense and lexical retrieval, tenant and knowledge-base isolation, role-based filtering, asynchronous document ingestion, and an MCP client for model-facing tool use. ### Key features: - **Pluggable Ingestion**: Supports Yuque, local Markdown connectors, and custom connectors. - **Hybrid Retrieval**: Combines vector similarity, BM25, and title relevance with reciprocal rank fusion. - **Agent-Ready API**: Provides versioned, tenant-isolated document ingestion, search, update, deletion, and ingestion-job endpoints. - **Stable Citations**: Returns document IDs, chunk IDs, source URLs, scores, metadata, and real character locations when available. - **Recoverable Writes**: Persists ingestion jobs and request snapshots before processing, with idempotency checks and stale-job recovery. - **Access Control**: Supports knowledge-base and document-level RBAC, public collections, user-token introspection, and service-token protection. - **MCP Integration**: Exposes remote retrieval and health checks as MCP tools without coupling the client to the platform internals. - **Evaluation Tooling**: Includes offline hit rate, recall, precision, MRR, and NDCG evaluation with baseline comparison. - **Deployment Flexibility**: Each runtime component has

### Overview: DeepCatalan is a project that aims to bring the Catalan language closer to deep learning models, specifically focusing on ULMFit. The repository is built with the goal of making possible to use DL models in Catalan, and it provides a comprehensive documentation of everything in Catalan. The project includes the training of ULMFit on Catalan Wikipedia articles, resulting in a model with a perplexity of 21.54 and a vocabulary of 30,656 words. The project also includes a dataset of 252,016 documents and over 98 million tokens. The repository is maintained by Adrià Cabeza and has been starred by 12 GitHub users. ### Key Features: - **ULMFit Model Training**: The project includes the training of ULMFit on Catalan Wikipedia articles, resulting in a model with a perplexity of 21.54 and a vocabulary of 30,656 words. - **Dataset**: The project provides a dataset of 252,016 documents and over 98 million tokens, which is used for training the ULMFit model. - **Classifying News Articles**: The project includes a dataset of 889 news articles scraped from various websites, and a classifying model that can categorize news articles into 5 categories: esports, economy, culture, technology, and health. ### Tech Stack: The project uses Jupyter Notebook as the primary language and is maintained by Adrià Cabeza. ### Why It Matters: The project aims

### Overview: A 180-billion-parameter mixture-of-experts model running on a single NVIDIA DGX Spark at 44 tokens per second, with quality cost measured and published. ### Key features: - **44 Tokens Per Second**: The model answers questions, codes, and prose at an impressive 44 tokens per second. - **65,536-Context Window**: The model operates with a 65,536-token context window, providing high-quality responses. - **Cross-Prompt Median**: The model's cross-prompt median performance is 43.34 tokens per second, with a target step of 55.27 milliseconds. ### Tech Stack: - **4-Bit NVFP4 + FP8 Hybrid**: The model is quantized using 4-bit NVFP4 and FP8 hybrid techniques. ### Why It Matters: - **Quality Cost Measurement**: The model's performance is measured and published, providing transparency and accountability. - **Experimental Model**: This is an experimental model that runs on a single machine, showcasing the potential of large language models on smaller hardware.

### Overview: PW-VQA is a project that facilitates the reproduction of results from the "Possible Worlds VQA" paper. The repository includes code based on implementations of RUBi and CF-VQA, aiming to provide a comprehensive environment for researchers to study causal visual question answering. ### Key features: - Reproducibility of results from the PW-VQA paper - Code based on RUBi and CF-VQA implementations - Access to pre-trained Skip-Thought encoder weights for text understanding ### Tech stack: Python ### Why it matters: PW-VQA provides a platform for researchers to study and reproduce the findings from the "Possible Worlds VQA" paper, focusing on causal visual question answering. The inclusion of pre-trained text encoder weights allows for easier experimentation and comparison with existing models.

Overview: This project fine tunes Mistral-7b using PEFT and LoRA on the Puffin dataset, focusing on multi-turn conversations between GPT-4 and real humans. Key features: - Fine-tuning Mistral-7b - PEFT (Parameter Efficient Fine-Tuning) - LoRA (Low-Rank Adaptation) - Puffin Dataset for multi-turn conversations Tech stack: Jupyter Notebook Why it matters: This project showcases advanced fine-tuning techniques for large language models, demonstrating their effectiveness on specific conversational data.

Overview: This repository implements the MeZO-SVRG algorithm for fine-tuning pre-trained language models, including fine-tuning on GLUE benchmark datasets like MNLI, QNLI, SST-2, and CoLA using the Hugging Face library. It leverages PyTorch and Pytorch Lightning, and supports fine-tuning on models such as distilbert-base-cased, roberta-large, gpt2-xl, and opt-2.7b. Key features: - MeZO-SVRG algorithm for fine-tuning language models - Supports GLUE benchmark datasets - Fine-tuning on models: distilbert-base-cased, roberta-large, gpt2-xl, opt-2.7b, opt-6.7b - GLUE tasks: MNLI, QNLI, SST-2, CoLA Tech stack: Python, PyTorch, Pytorch Lightning Why it matters: This repository provides a practical implementation of variance-reduced zero-order methods for fine-tuning language models, offering a more efficient approach compared to first-order methods.

### Overview: CoGeLoT is a comprehensive framework designed to investigate the resilience of embodied multimodal models. It explores whether these models are plausibly resilient by examining the role of instruction variety and task difficulty. ### Key Features: - **Model Architectures and Provided Checkpoints:** CoGeLoT offers a variety of model architectures and checkpoints for evaluation. Users can easily switch between different models and run experiments using the provided checkpoints. - **Datasets:** The framework utilizes datasets hosted on Hugging Face, ensuring easy access and reproducibility. ### Tech Stack: - Python 3.11 - PyTorch - Lightning ### Why it Matters: CoGeLoT provides a robust and comprehensive approach to evaluating the resilience of multimodal models, making it a valuable tool for researchers and developers in the field of robotics and embodied AI. --- Repository: amitkparekh/CoGeLoT GitHub Stars: 12 GitHub Description: A comprehensive framework to explore whether embodied multimodal models are plausibly resilient

### Overview: Haru is a GPU Hardware Abstraction Layer (HAL) designed for LLM inference fleets, inspired by the role the HAL plays in an operating system. It extracts GPU lifecycle management logic into an independent layer, allowing product control planes to focus on users, catalogs, and metadata while consuming Haru through a small HTTP API. ### Key Features: - **GPU Lifecycle Management**: Haru manages GPU provisioning, supervision, and failover for LLM inference fleets. - **Active/Standby Architecture**: The active domain serves OpenAI-compatible inference traffic, while the standby domain keeps the same model runtimes resident in level 1 sleep mode, freeing standby GPUs for preemptible LoRA training. - **Hot Failover**: Haru promotes the standby domain when the active domain fails, ensuring zero idle GPUs and efficient use of resources. ### Tech Stack: TypeScript ### Why it Matters: Haru simplifies the management of LLM inference fleets by separating GPU lifecycle management from the product control plane, allowing for a more focused and efficient product design.

Overview: This repository contains tested image prompt recipes for Seedream 5.0 Pro, including 8 recipes with their corresponding renders, aspect ratios, and the model ID and cost per image. Key features: - 8 Seedream 5.0 Pro prompts - Example renders, aspect ratios, model ID, and cost per image - Easy-to-follow cheat sheet for prompt structure Tech stack: Python Why it matters: This repository provides practical, tested image prompt recipes for Seedream 5.0 Pro, ensuring high-quality renders and efficient cost management. Repository: ayekoa913g/seedream-5-0-pro-prompts-id Primary language: Python GitHub stars: 11

### Overview: Seedance 2.0 API (`seedance-2.0`) is a comprehensive OpenAI-compatible API endpoint designed for generating high-quality images, videos, and text-to-video content. It offers a cost-effective solution without the need for a monthly subscription, making it ideal for developers and content creators looking to integrate AI capabilities into their projects. ### Key Features: - **Model ID and Pricing**: The API supports a range of models, including text-to-video and image-to-video, with customizable resolutions and durations. - **Pricing Model**: The pricing is based on the number of generated seconds, with a default rate of $0.04 per second. Users can choose from 480P, 720P, and 1080P resolutions. - **Asynchronous Processing**: The API is designed to be highly efficient, allowing users to start generating content without waiting for the entire process to complete. - **Model Compatibility**: Seedance 2.0 is compatible with multiple models, including text-to-video and image-to-video, providing a versatile solution for various creative projects. ### Tech Stack: - **Python**: The API is implemented using Python, ensuring compatibility and ease of integration with existing projects. ### Why It Matters: - **Cost-Effective**: Seedance 2.0 offers a cost-effective alternative to more expensive solutions, making it accessible for developers on a budget. - **Ease of Use**: The API is designed to be user-friendly, with clear documentation and examples provided to help developers get started quickly. - **High-Quality

### Overview: Seedance 2.0 Prompts: 8 Tested Recipes with Outputs Seedance 2.0 Prompts offers tested image prompt recipes, aspect ratios, the model ID (`seedance-2.0`), and the cost per image. This repository provides example renders, duration, and the call costs for each recipe. It is compatible with OpenAI's API and costs $0.066 per second for 480P (720P $0.216, 1080P $0.3849) without a monthly subscription. ### Key Features: - **Model ID:** `seedance-2.0` - **Recipes:** 4 (available in `data/prompts.json` and `PROMPTS.md`) - **Effective Price:** $0.066 per second (estimated cost of the whole set: $0.264) ### Tech Stack: - Python ### Why It Matters: - **A Single Endpoint:** Access Seedance 2.0 (`seedance-2.0`) and over 300 other models with a single API key. - **Flexible Pricing:** Pay only for what you use, starting at $1 per call. - **Real-Time Pricing:** The cost is reflected in the response, and failed calls are refunded. - **Asynchronous by Design:** Use the `task_id` to track and retry tasks. ### Also Known As: - Dreamina ### What Separates a Usable Recipe from a Wasted Call:

### Overview: Grok Imagine 1.5 Edit API is a powerful tool for image editing, offering a single endpoint that can generate images for any OpenAI-compatible model. It supports various resolutions and pricing models, making it ideal for teams looking to integrate image editing into their workflows without the need for a monthly subscription. ### Key Features: - **One Key, Entire Catalog**: The same `https://api.apimart.ai/v1` base URL and `Authorization` header can reach Grok Imagine 1.5 Edit and over 300 other image, video, and language models. - **$1 Minimum, Pay As You Go**: No subscription required, and there's no free quota to burn through. The price is reported in the response for each call. - **Billing Model**: The charge comes back in the response, so a spend number is read per call instead of guessed at month end. - **Async by Design**: Submit a task, poll the `GET /v1/tasks/{id}` endpoint for results, and the API handles batching and retries automatically. ### Tech Stack: - **Python**: The primary language used for development and integration. ### Why It Matters: - **One Key, Entire Catalog**: The same base URL and header can reach multiple models, simplifying integration. - **$1 Minimum, Pay As You Go**: No subscription required, and there's no free quota to burn through, making it flexible and cost-effective. - **Billing Model**: The charge is reported in the response, providing transparency and ease of tracking costs. - **Async