AI Trending

Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 11 stars

kwmw8z1tvj5/seedance-1-0-pro-fast-prompts-ja

シードダンス / シーダンス — Seedance 1.0 Pro Fast prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`doubao-seedance-1-0-pro-fast`) and the cost per image.

### Overview: Seedance 1.0 Pro Fast Prompts is a comprehensive set of tested image prompt recipes designed for developers. It includes eight recipes with detailed outputs, aspect ratios, and the model ID (`doubao-seedance-1-0-pro-fast`) and cost per image. ### Key Features: - **8 Recipes with Outputs**: Detailed examples and cost information for each recipe. - **Model ID**: `doubao-seedance-1-0-pro-fast` for seamless integration with the Seedance API. - **Cost Per Image**: $0.0079 per second, billed by resolution tier. - **OpenAI-Compatible Endpoint**: Supports OpenAI-compatible endpoints, no need for a monthly plan. ### Tech Stack: - Python ### Why It Matters: - **Key Features**: The recipes are tested and optimized for high-quality image generation. - **Cost Efficiency**: Offers a low-cost solution with a flexible pricing model. - **Model Integration**: Seamless integration with the Seedance API, reducing client-side complexity. - **Prompt Structure**: Provides a cheat-sheet for crafting effective prompts, ensuring better results. ### Repository: - **GitHub Repository**: kwmw8z1tvj5/seedance-1-0-pro-fast-prompts-ja - **Stars**: 7 - **Description**: Seedance 1.0 Pro Fast Prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`doubao-seedance-1-0-pro-fast`) and the cost per image. ### Primary Language
Python 11 stars

l-ca0/midjourney-api-pt

Midjourney API (`midjourney`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Midjourney API, also known as `midjourney`, is a programmable API that exposes Midjourney as a text-to-image engine, offering a range of models and pricing options. Key features: - Supports various image, video, and language models - Charges per delivered image, with tiered pricing based on resolution - Offers default, blend, fast, and turbo queues for different performance levels - Supports both asynchronous and synchronous API calls Tech stack: Python Why it matters: Midjourney API provides a cost-effective and efficient way to generate high-quality images and videos using the industry-leading Midjourney model, making it ideal for developers and creatives looking to integrate AI-powered image generation into their projects. GitHub stars: 8
Python 11 stars

lh0x00/docsifer

Docsifer is a powerful tool for converting various data formats into Markdown for applications such as indexing, text analysis, and more. It supports PDF, PowerPoint, Word, Excel, Images, Audio, HTML, and other text-based formats, and leverages LLMs to enhance performance.

Overview: Docsifer is a powerful tool designed to convert various document formats into clean, LLM-ready Markdown, leveraging LLMs for enhanced performance. Key features: - Supports PDF, Word, PowerPoint, Excel, HTML, audio, images, CSV, JSON, ZIP, and other text-based formats. - Optional LLM for OCR, transcription, and structured layout extraction. - Production-grade features including bounded concurrency, per-IP fairness, memory watchdog, disk cleanup, and circuit breaker. - Hardened against SSRF, path traversal, body-size limits, and security headers. - Observability with JSON logs and `/v1/healthz`, `/v1/readyz`, `/v1/stats` endpoints. - Privacy-first processing in memory and immediate discard. Tech stack: Python. Why it matters: Docsifer simplifies the conversion of documents into Markdown, making them more suitable for applications like indexing and text analysis, while enhancing performance with the use of LLMs. Repository: lh0x00/docsifer Primary language: Python GitHub stars: 10 GitHub description: Docsifer is a powerful tool for converting various data formats into Markdown for applications such as indexing, text analysis, and more. It supports PDF, Word, PowerPoint, Excel, HTML, audio, images, CSV, JSON, ZIP, and other text-based formats, and leverages LLMs to enhance performance.
C++ 11 stars

loopai-hq/pulsar

Pulsar (formerly fastkernel): the fastest inference engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon

**Overview:** Pulsar is a cutting-edge inference engine designed specifically for Qwen3.8-27B and Qwen3.6-35B-A3B models, optimized for Apple Silicon processors. It achieves remarkable performance, writing code at 156 tokens per second and editing code at 347 tokens per second, making it the fastest engine for these models on the Apple Silicon platform. **Key features:** - Pulsar is the fastest engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon, surpassing Lithos-Metal 0.1.2 by 1.13 times and Splash 1.3.0 by 0.13 times. - It writes code at 156 tokens per second and edits code at 347 tokens per second, significantly faster than Lithos-Metal 0.1.2. - Pulsar supports the full model checking every token, ensuring accuracy. **Tech stack:** C++ with Python bindings. **Why it matters:** Pulsar is essential for developers who need the fastest possible performance when working with Qwen models on Apple Silicon, providing a significant performance boost in both coding and editing tasks. **GitHub stars:** 10 **Primary language:** C++ **GitHub description:** Pulsar (formerly fastkernel): the fastest inference engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon
HTML 11 stars

lqhl/vectordb-comparison

Overview: This repository offers a detailed comparison of various vector databases, focusing on their unique features, capabilities, and performance metrics. It aims to help users and developers make informed decisions when choosing a vector database for their specific needs. Key features: - MyScale - Pinecone - Qdrant - Weaviate - Milvus - Chroma - Postgres with pgvector - OpenSearch Tech stack: HTML Why it matters: This repository serves as a valuable resource for developers and users looking to compare and choose the best vector database for their applications, ensuring they make the most informed decision based on the specific needs and requirements of their projects.
Python 11 stars

lrj-zpge/grok-imagine-image-2-0-api-de

Grok Imagine Image 2.0 API (`grok-imagine-image-2.0`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Grok Imagine Image 2.0 API provides access to high-fidelity image generation via an OpenAI-compatible endpoint, offering flexible pricing and model compatibility. Key features: - Supports multiple image, video, and speech models - Pricing per image by resolution and quality - Model id and routes - Pricing details and effective price - Request parameters for prompt, size, resolution, and more Tech stack: Python Why it matters: This API allows developers to create high-fidelity images with flexible pricing and model compatibility, making it ideal for creative projects and applications that require precise image generation.
Python 11 stars

lucataco/DeepSeek-V4.1-Flash-2x-DGX-Spark

DeepSeek-V4.1-Flash-1017 on two NVIDIA DGX Sparks (GB10) with TensorFold TP2: scripts, chat-template fix, speed and quality results

Overview: This repository contains scripts and fixes for DeepSeek-V4.1-Flash-1017, a version optimized for two NVIDIA DGX Spark nodes with TensorFold TP2. The scripts include a chat-template fix and improvements in speed and quality. The project is based on the DeepSeek-V4.1 engine with 1M-token context, 4 concurrent sessions, and speculative decoding. Key features: - Scripts for reproducing the results - Fix for the 1017 release - Improved speed and quality - TensorFold engine with RoCE and TensorFold TP2 - 1M-token context and 4 concurrent sessions - DGX Spark nodes with 128 GB unified memory Tech stack: Python Why it matters: This project provides a reliable and efficient way to deploy DeepSeek-V4.1-Flash-1017 on two DGX Spark nodes, improving both speed and quality of the model's output. The scripts and fixes ensure that the model can be easily reproduced and tested, making it valuable for developers and researchers. GitHub stars: 11 Repository: lucataco/DeepSeek-V4.1-Flash-2x-DGX-Spark
Python 11 stars

mahimairaja/locallens

Search your files by talking to them - 100% offline

Overview: LocalLens is a powerful file search and retrieval tool that allows users to search their files by meaning or exact match, offering a hybrid retrieval system. It supports RAG Q&A, voice input and playback, and is fully offline, making it perfect for remote and offline work environments. Key features: - Semantic + keyword search - RAG Q&A - Voice input + playback - CLI + web app - Watch mode - Namespaces - Plugin extractors Tech stack: Python Why it matters: LocalLens simplifies file management by providing a seamless, offline-first solution. It integrates seamlessly with AI agents like Claude Code, Claude Desktop, and others, making it a versatile tool for developers and professionals alike. Repository: mahimairaja/locallens GitHub stars: 11 GitHub description: Search your files by talking to them - 100% offline GitHub logo: assets/logo.svg GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description: Search your files by talking to them - 100% offline GitHub description
Julia 11 stars

mantzaris/WunDeeDB.jl

Your just-works / zero-config / embedded / WAL: vector database

Overview: WunDeeDB.jl is a Julia package designed for storing and querying vector embeddings in a SQLite database, supporting various numerical types and approximate nearest neighbor indices like HNSW and LM-DiskANN. Key features: - Supports `Float16`, `Float32`, `Float64`, and integer types - Provides a linear fallback for distance queries if no ANN method is enabled - SQLite backend for efficient storage and retrieval Tech stack: Julia Why it matters: WunDeeDB.jl offers a simple and zero-config solution for embedding storage and retrieval, making it suitable for applications requiring fast and accurate vector database operations. Repository: mantzaris/WunDeeDB.jl GitHub stars: 11 GitHub description: Your just-works / zero-config / embedded / WAL: vector database Documentation: https://mantzaris.github.io/WunDeeDB.jl/ License: MIT Build Status: https://github.com/mantzaris/WunDeeDB.jl/actions DOI: https://doi.org/10.21105/joss.08033
Python 11 stars

memseekai/memseek

Memseek is an open source declarative context engine for AI agents. Define agent memory in versioned YAML — like Terraform for infrastructure. Memseek learns from everything you record and gives your agents cited, current, budget-cut context from a Postgres you run.

### Overview: Memseek is an open-source declarative context engine for AI agents, designed to turn versioned YAML records into up-to-date, evidence-backed context for your agents. ### Key features: - **Versioned YAML Memory**: Define agent memory in YAML, versioned and immutable. - **Context Compilation**: Converts records into context packages, ensuring current, cited, and budgeted context from Postgres. - **Immutable Records**: No retrieval pipeline, opaque notebooks. Every fact is dated, cited, and traceable. - **Immutable Events**: Records are immutable, ensuring the integrity of the context. - **Immutable State**: State is versioned, preventing accidental changes. - **Immutable Schema**: All records and state are schema-validated, ensuring consistency. - **Immutable Views**: Views are immutable, preventing unauthorized modifications. - **Immutable Artifacts**: Artifacts are immutable, preserving the integrity of the context. - **Immutable Sessions**: Sessions are immutable, preventing unauthorized access to the database. - **Immutable Commands**: Commands are immutable, ensuring the integrity of the context. - **Immutable Files**: Files are immutable, preventing unauthorized access to the database. ### Tech stack: Python ### Why it matters: Memseek turns versioned records into up-to-date, evidence-backed context for AI agents, ensuring current, cited, and budgeted context from Postgres. It eliminates the need for opaque notebooks, retrieval pipelines, and opaque derivations, making it a powerful tool for managing context and ensuring the integrity of AI agents.
Python 11 stars

mkamranr/airllm-openai-server

OpenAI-compatible API server wrapping AirLLM (github.com/lyogavin/airllm) — run large LLMs on small GPUs, CPU, Apple silicon or Docker

Overview: This project is an OpenAI-compatible API server that wraps AirLLM, allowing you to run large language models on small GPUs, CPUs, Apple silicon, or Docker. It provides a simple way to integrate AirLLM models into existing OpenAI-compatible clients. Key features: - Run large LLMs on small GPUs, CPUs, Apple silicon, or Docker - Use the same API endpoints as OpenAI clients - Stream inference work, making large models run efficiently on small resources Tech stack: Python Why it matters: This project simplifies the integration of AirLLM models into existing OpenAI-compatible clients, making it easier to deploy large language models on various hardware configurations without needing to rewrite the entire API. Repository: mkamranr/airllm-openai-server GitHub stars: 9 GitHub description: OpenAI-compatible API server wrapping AirLLM (github.com/lyogavin/airllm) — run large LLMs on small GPUs, CPU, Apple silicon or Docker
Jupyter Notebook 11 stars

mpuig/gpt2-fine-tuning

Fine-tune GPT2 to generate fake job experiences

### Overview: Fine-tune GPT-2 model to generate fake job experience descriptions. ### Key features: - Fine-tuning GPT-2 language model for generating job experience descriptions. - Dataset of real job experiences from LinkedIn. - Goal: Create a model that can generate complete, relevant sentences related to job titles. ### Tech stack: Jupyter Notebook ### Why it matters: This repository serves as a valuable learning opportunity for understanding the fine-tuning process of a language model. By experimenting with generating job experience descriptions, the developer can gain insights into GPT-2's capabilities and limitations in generating realistic and relevant content.
Jupyter Notebook 11 stars

nanom/llm_adaptation_2nlp_workshop

Taller: Modelos de lenguaje a tu medida. 2º Workshop Argentino de NLP, Córdoba 2023

Overview: This project is a Jupyter Notebook-based workshop focused on adapting language models to individual needs. Key features: - Interactive Jupyter Notebook environment - Hands-on experience with language model adaptation - Practical application of NLP techniques Tech stack: Jupyter Notebook Why it matters: This workshop provides developers with a practical way to tailor language models to specific applications, enhancing their utility and effectiveness.
Kotlin 11 stars

nareshis21/Truelarge-RT

Android inference engine running 20B+ parameter LLMs on 4GB-8GB RAM devices. Features proprietary Layer-by-Layer (LBL) streaming, zero-copy mmap loading, and native C++/Kotlin architecture.

### Overview: TrueLarge-RT is a high-performance inference engine that enables 32B+ parameter LLMs to run on consumer Android devices without crashing. It uses a proprietary Deep-Pipelined Layer-by-Layer (LBL) engine to stream weights from storage, offering infinite scalability limited only by your disk size. ### Key Features: - **No Quantization Required**: Run full precision (FP16) or high-quality (Q4_K_M) GGUF models. - **Native Performance**: Built on `llama.cpp` with custom ARM NEON/DotProd kernels. - **Resilient Model Manager**: Integrated background downloader for massive model files. - **Professional Telemetry**: Real-time graphs for TTFT (ms), RAM (MB), and CPU (GHz). - **Universal Support**: Works with Llama 3, Qwen 2.5, Mistral, Phi-3, and more. - **Mixture-of-Experts (MoE)**: Support is actively being researched. ### Tech Stack: Kotlin ### Why it Matters: TrueLarge-RT successfully runs Llama-3.3-70B-Instruct-Q2_XS on a Realme 2 Pro (4GB RAM) with only ~1400MB available RAM at 0.01 TPS, demonstrating absolute scalability on legacy hardware constrained by UFS 2.1 storage speeds.
Rust 11 stars

neilthomass/kowari

Lightweight vector database in Rust with HNSW, LSH, and custom encoding

Overview: Kowari is a lightweight Rust vector database that supports multiple search algorithms, flexible storage backends, high performance, and privacy-first features. Key features: - Multiple Search Algorithms: Brute-force search, Locality-Sensitive Hashing (LSH), and Hierarchical Navigable Small World (HNSW) for approximate search. - Flexible Storage Backends: In-memory storage, custom `.kwi` binary format, and SQLite integration for metadata and system information. - High Performance: Written in Rust for maximum speed and safety, optimized vector operations with SIMD support, and memory-efficient data structures. - Privacy First: 100% local storage, no cloud dependencies or external services, complete control over vector data. Tech stack: Rust. Why it matters: Kowari is a fast, local vector database that can store, search, and manage vectors efficiently, making it ideal for applications that require high performance and privacy. Repository: neilthomass/kowari GitHub stars: 11 GitHub description: Lightweight vector database in Rust with HNSW, LSH, and custom encoding.
Python 11 stars

neshani/illumination_pipeline

An automated pipeline to convert e-books into beautifully illustrated scenes using local AI backends ComfyUI and SD Forge.

Overview: An automated pipeline for converting e-books into beautifully illustrated scenes using local AI backends ComfyUI and SD Forge. Key features: - EPUB to Text: Extracts clean, readable text from `.epub` files. - AI Prompt Generation: Connects to an OpenAI-compatible local LLM server for context-aware image prompts. - Smart Chunking: Groups paragraphs into larger, configurable chunks. - Dual Backend Support: Works with both SD Forge and ComfyUI. - Batch Image Generation: Processes the entire list of prompts. - Batch Upscaling: Upscales all generated images. - Config-Driven Workflows: Settings controlled via JSON configuration files. - Interactive CLI: User-friendly interface for easy navigation. Tech stack: Python Why it matters: This tool automates the tedious process of converting e-books into illustrations, making the creation of AI art more accessible and efficient. Repository: neshani/illumination_pipeline GitHub stars: 11
Python 11 stars

nicolay-r/THOR-ECAC

The official fork of THoR Chain-of-Thought framework, enhanced and adapted for Emotion Cause Analysis (ECAC-2024)

### Overview: THOR-ECAC is an official fork of the THoR Chain-of-Thought framework, adapted for Emotion Cause Analysis (ECAC-2024). This repository contains the PyTorch reforged and forked version of the original THoR framework, enhanced and tailored for the SemEval-2024 task. ### Key features: - Enhanced and adapted for ECAC-2024 - PyTorch implementation - Supports batched mode for faster inference - Incorporates Flan-T5 model for reasoning on emotion cause in conversations ### Tech stack: Python ### Why it matters: This repository provides a practical tool for researchers and developers interested in emotion cause analysis and chain-of-thought reasoning, offering a robust framework for the SemEval-2024 task.
Python 11 stars

ozcanmiraay/vidur-energy

Energy and carbon-aware extension of Vidur for sustainable LLM inference simulation and analysis 💡🌳

### Overview: Vidur-Energy is an enhanced version of Vidur, a high-fidelity LLM inference system simulator, that includes additional energy tracking capabilities. This extension enables more sustainable AI inference optimizations by monitoring power draw, estimating energy consumption, and calculating carbon emissions. ### Key Features: - **Power Tracking**: Extracts GPU power draw at different utilization levels. - **Energy Consumption Estimation**: Tracks energy usage across inference workloads. - **Carbon Footprint Estimation**: Uses grid carbon intensity data to estimate inference-related emissions. - **Modular Energy Tracking**: A more configurable and extensible approach to tracking energy metrics. - **Full Compatibility with Vidur**: All existing simulation capabilities remain unchanged. ### Tech Stack: Python ### Why It Matters: This repository provides a comprehensive tool for quantifying the energy consumption and carbon emissions of large language model (LLM) inference, helping developers optimize their AI inference processes for sustainability.
Python 11 stars

pacificrm/RAG-Architecture

Multimodal Document Processing RAG with LangChain

Overview: This project is a Streamlit application designed to process multimodal documents and query a Milvus database. It utilizes LangChain and EasyOCR for text extraction from various file types, and HuggingFaceEmbeddings for vector representation generation. The system supports multiple file types and offers a natural language query interface for AI-driven responses. Key features: - Upload file processing: Extracts text from audio, video, images, documents, and more. - Milvus integration: Stores and queries document embeddings. - Query interface: Implements a Retrieval-Augmented Generation (RAG) pipeline. Tech stack: Python, Streamlit, LangChain, EasyOCR, HuggingFaceEmbeddings, Milvus. Why it matters: This project provides a comprehensive solution for processing and querying multimodal documents, leveraging cutting-edge AI and database technologies to enhance the efficiency and accuracy of information retrieval and processing. Repository: pacificrm/RAG-Architecture GitHub stars: 11 GitHub description: Multimodal Document Processing RAG with LangChain GitHub URL: https://github.com/pacificrm/RAG-Architecture
Python 11 stars

pcrgss12wsx7w9/doubao-seedance-1-0-pro-api-id

Seedance 1.0 Pro API (`seedance-1.0-pro`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: Seedance 1.0 Pro API (`seedance-1.0-pro`) is a fast AI video model designed for high-quality video generation and rapid previews. It offers a single API key that covers 300+ models including images, videos, and multiple languages, without the need for monthly subscriptions or additional fees. ### Key features: - **Single Key for All Catalogs**: A single API key that covers all models in the Seedance catalog, including 300+ image and video models, and multiple languages. - **Pricing by Second**: The API is billed based on the number of generated seconds, with a tiered pricing model. - **No Monthly Subscription**: No need for a monthly subscription; pay only for usage. - **Real-Time Billing**: Each task reports the exact cost, and tasks with failed attempts are refunded. - **Asynchronous Processing**: Tasks start processing immediately, and you can track their status using `GET /v1/tasks/{task_id}`. ### Tech Stack: - Python ### Why it Matters: - **Unified Key for All Models**: A single key covers all models, eliminating the need to manage multiple keys. - **Flexible Pricing**: No need for monthly subscriptions, and pay only for usage. - **Real-Time Billing**: Tasks report costs accurately, and failed attempts are refunded. - **Asynchronous Processing**: Tasks start processing immediately, making it easy to track their status. ### Repository: - **Repository**: pcrgss12wsx7w9/doubao-seedance-1-0