Projects

Overview: This project provides a comprehensive set of tools and resources for developers working with the OpenAI-compatible API endpoint `flux-2-flex`, allowing them to generate and edit images with adjustable steps and guidance for fine control and typography. Key features: - **Model ID and Routes**: Supports various image generation and editing capabilities with adjustable steps and guidance for fine control and typography. - **Pricing**: Offers flexible pricing with a per-delivered-image model, starting at $0.04 per image. - **Model Page**: Provides detailed information on the available models and their pricing tiers. - **Request Parameters**: Accepts a variety of parameters including `model`, `prompt`, `size`, `resolution`, `n`, and `image_urls`. Tech stack: Python Why it matters: This project is essential for developers looking to integrate the OpenAI-compatible API endpoint `flux-2-flex` into their applications, offering a flexible and efficient way to generate and edit images with adjustable steps and guidance for fine control and typography. Repository: bsl2e86y9/flux-2-flex-api-ko Primary language: Python GitHub stars: 12 GitHub description: 플럭스 — FLUX.2 Flex API (`flux-2-flex`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: This repository provides a comprehensive solution for performing Supervised Fine-Tuning (SFT) on the Qwen2.5-Coder-1.5B-Instruct model using the nvidia/OpenCodeReasoning dataset. It includes all necessary components for setting up and running the model. Key features: - All required dependencies are included in the `requirements.txt` file. - The model is fine-tuned on the Qwen Models using the nvidia/OpenCodeReasoning dataset. - The project structure is well-defined with clear documentation and a README. Tech stack: Python Why it matters: This repository offers a practical and efficient way to fine-tune models, making it valuable for developers looking to enhance their models with new data and techniques. GitHub stars: 12 Repository: bunyaminergen/Qwen2.5-Coder-1.5B-Instruct-Reasoning

Overview: Qwen-Image 2.0 API offers powerful AI image generation capabilities, providing high-quality images from text prompts or transforming existing images. It supports over 300 models, including Qwen, Pro models, and more. Key features: - Supports Qwen and Pro models - Generates high-quality images from text prompts - Transforms existing images with Qwen standard and Pro models - Compatible with OpenAI-compatible endpoint - Pricing model: $0.02 per image, $1 minimum, pay-as-you-go - Asynchronous processing, task results available via `GET /v1/tasks/{task_id}` Tech stack: Python Why it matters: Qwen-Image 2.0 API simplifies the process of generating high-quality images, making it accessible for developers and businesses looking to integrate AI-powered image generation into their projects.

### Overview: This repository contains a set of tested image prompt recipes for the Qwen-Image 3.0 Pro model, including examples, aspect ratios, and the cost per image. ### Key Features: - **8 Recipes**: Eight different prompts for the Qwen-Image 3.0 Pro model. - **Example Renders**: Each recipe includes an example render, the aspect ratio, and the cost of the call on this route. - **Prompt Structure**: A cheat-sheet for structuring prompts, including the subject, light, framing, setting, and finish. ### Tech Stack: Python ### Why It Matters: These recipes are designed to help developers create high-quality images efficiently, ensuring that the model is used effectively and at a reasonable cost.

### Overview: Dual-Radeon vLLM is a project that benchmarks and optimizes language models on consumer AMD Radeon graphics cards, specifically focusing on the RX 7900 XT, 7900XTX, and RX 7900XT. The project aims to diagnose and fix the RCCL hostcall / PCIe atomics crash, which is a root cause for the RCCL failure on consumer Radeon cards. The benchmarks are performed on 13 different machine configurations, including a virtual machine with no P2P and cross-die PCIe 3.0, and the results are used to measure performance improvements and identify bottlenecks. ### Key Features: - **RCCL Bug Fix**: A 57-line reproducer for the RCCL bug that causes the `--tensor-parallel-size 2` option to fail on consumer Radeon cards, specifically due to missing PCIe AtomicOps. - **Performance Benchmarks**: Seven model architectures on fourteen machine configurations, including two consumer Radeons together and apart, an A100 80G and 40G, an L4 24G, a Tesla T4 16G, and a rented H100 (one, two, and four of them), B300, and RTX PRO 6000 (one and two), and a rented MI300X. - **Data Analysis**: Raw per-request records, runners, and analysis scripts that do not require a GPU, including information on card clocks, power, and temperature, and

Overview: Qwen-Image 3.0 Pro is a high-fidelity image generation API that offers 1K and 2K output options at a cost of $0.0285712 per image. Key features: - Supports OpenAI-compatible endpoint - Access to 300+ models via a single base URL and `Authorization` header - No need for monthly subscriptions or pre-paid plans - Charges based on response cost - Designed for non-blocking requests with simple polling Tech stack: Python Why it matters: Qwen-Image 3.0 Pro provides a cost-effective way to access high-fidelity image generation capabilities without the need for a monthly subscription or pre-paid plans. The API allows direct access to the OpenAI-compatible endpoint, making it ideal for developers looking to integrate advanced image generation into their projects without additional setup or maintenance.

### Overview: 在阿里巴巴千问杯多模态推理大赛中，我们使用Qwen3.5-0.8B多模态模型构建了一套教师蒸馏SFT与小剂量GRPO强化学习相结合的训练流水线。该方案覆盖了数值推理、计数、空间推理、物理常识、感知、关联六维能力。核心产出包括奖励-评测口径逐字对齐、同权重ZIP CRC消融、统计化提交额度决策以及0.8B小模型RL剂量-遗忘权衡的系统刻画。最终，该方案在初赛中获得了41.8分，决赛中取得了20.2分，位列决赛第9/20名，总分逆风翻盘至第3名，获得了2万元奖金。 ### Key features: 1. **教师蒸馏SFT**: 利用Qwen3.6-27B教师模型对113,159张候选图像进行蒸馏，生成113,159条样例。 2. **GRPO小剂量强化学习**: 小剂量GRPO算法，30步，学习率1e-6，β0.2，有效batch8，总采样240个样本。 3. **奖励-评测口径逐字对齐**: 答案提取器与官方评测脚本逐字一致，无reward hacking空间。 4. **同权重ZIP CRC消融**: 通过消融实验验证了奖励-评测口径的准确性

### Overview: Imagina 4.0 prompts is a collection of 8 tested recipes for generating high-quality image prompts. Each recipe includes an example render, the aspect ratio, and the cost on the `imagen-4.0-apimart` route. This resource is useful for developers looking to optimize their image generation processes. ### Key Features: - **8 Recipes**: Eight different prompts tested with example renders, aspect ratios, and the cost on the `imagen-4.0-apimart` route. - **Model ID**: `imagen-4.0-apimart` for generating images. - **Cost**: $0.04 per image, estimated cost for the entire set: $0.32. - **Prompt Structure**: A cheat-sheet for structuring prompts, including subject, light, framing, setting, and finish. ### Tech Stack: Python ### Why It Matters: - **Cost Efficiency**: The cost is $0.04 per image, making it an affordable option. - **Model Compatibility**: The same base URL and API key can be used for multiple models, including Imagen 4.0. - **Real-time Pricing**: The cost is displayed in the response, and each completed task is refunded for failed attempts. - **Asynchronous Processing**: The API is designed to be asynchronous, allowing for efficient processing of multiple tasks. ### Repository: - **Repository**: `cdi8dv7g2d4/imagen-4-0-prompts-fr` - **GitHub Stars**: 12 - **

Overview: Seedance 2.5 API, also known as Seedance, is an OpenAI-compatible endpoint providing video generation services with cinematic motion and per-second billing by resolution. It offers a single key for access to multiple models and eliminates the need for monthly fees. Key features: - Access to 300+ models including video, image, and language, with a single base URL and authorization header. - Billing is per generated second by resolution, with a minimum cost of $1 per task. - Summaries are returned in the response, and tasks with failed status are reported back to the balance. - Tasks are processed asynchronously by default, with task IDs and status polling for real-time updates. Tech stack: Python. Why it matters: Seedance 2.5 API simplifies access to high-quality video generation services, making it an attractive option for developers looking to integrate AI-driven video creation into their applications without the hassle of traditional licensing models.

### Overview: Qwen-Image 3.0 Pro API is a powerful tool for generating high-fidelity images using the Qwen model, offering 1K/2K output resolution at a cost of $0.0285712 per image. It supports OpenAI-compatible pricing and is billed per delivered image, making it ideal for artists, designers, and anyone looking to create detailed and realistic images. ### Key Features: - **Model ID and Routes**: The API uses a model id of `qwen-image-3.0-pro` and supports different output resolutions (1K and 2K). - **Pricing**: The pricing is $0.0285712 per image, with the option to choose between 1K and 2K resolution. - **Model Page**: Detailed information on the model can be found on the [model page](https://go.apimart.ai/k-dbc985). - **API Key**: Requires an API key to access the API. - **Real-time Pricing**: Prices are updated in real-time and can be checked on the [pricing page](https://go.apimart.ai/k-4091fd) and `data/model.json`. - **Task Status**: Tasks can be monitored using the `GET /v1/tasks/{task_id}` endpoint. - **Model Parameters**: The prompt, size, resolution, and number of images can be customized to suit specific needs. ### Tech Stack: - Python ### Why It Matters: - **Key Feature**: A single key

Overview: GPT Image 1 API (`gpt-image-1`) is an OpenAI-compatible image generation model provided by APIMart, offering token-based pricing and a single endpoint for both text and image inputs. Key features: - Supports OpenAI-compatible API - Provides a single endpoint for text and image inputs - Offers token-based pricing - Supports various image resolution tiers - Supports both text and image prompts - Supports multiple image inputs per task - Tracks and reports costs and credits Tech stack: Python Why it matters: This API is useful for developers looking to integrate image generation capabilities into their applications, offering a cost-effective and OpenAI-compatible solution for generating images based on prompts or editing existing images. Repository: cpbkfsk6aj/gpt-image-1-api-es Primary language: Python GitHub stars: 12 GitHub description: ChatGPT Imágenes — GPT Image 1 API (`gpt-image-1`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: vectile: your private library A local RAG (Retrieval Augmented Generation) app for Windows, macOS, and Linux. vectile indexes your notes, books, and code into one SQLite database with hybrid vector + full-text search (Reciprocal Rank Fusion), making it easy to find anything with local hybrid search and expose through MCP. It connects with your Obsidian vault, projects, Calibre library, and code repos, and runs on your machine. ### Key features: - **Private:** indexing, parsing, embedding, search, and OCR run inside the app. Your documents and their text are never uploaded. See [network activity](#network-activity) for the three requests the app does make. - **Search by meaning:** a blue-green deploy note matches “how do we ship changes safely” without using those words. - **Index what you have:** Obsidian vaults, document folders (Markdown, PDF, DOCX, HTML, TXT, CSV, JSON, YAML, XML, SQL, shell scripts, XLSX, PPTX, Jupyter notebooks, EPUB), Calibre libraries, and code repositories with commit history. - **Bring your own model:** import or download a `.gguf` from Settings. - **Scanned PDFs:** install the Tesseract OCR plugin in Settings; photo pages become searchable. - **Manage the library:** browse files and chunks, delete a source, selected chunks, or a whole library. - **MCP server:** expose search, read, and collection tools to any local MCP client

Overview: This repository offers tested Seedream 4.5 prompts, including 8 recipes with example renders, aspect ratios, the model ID (`seedream-4-5`), and the cost per image. The prompts are available through the APIMart API, providing a cost-effective alternative to OpenAI's models. Key features: - 8 Seedream 4.5 prompts - Example renders, aspect ratios, and costs - Model ID: `seedream-4-5` - Effective price: $0.026 per image, estimated cost of the whole set: $0.208 Tech stack: Python Why it matters: Seedream 4.5 prompts offer a cost-effective alternative to OpenAI's models, providing high-quality image outputs without the need for a subscription or monthly plan.

### Overview: FLUX 2 Max Prompts: 8 Tested Recipes With Outputs ### Key Features: - **Model:** `flux-2-max` - **Prompts:** 8 (machine-readable in `data/prompts.json`, table in `PROMPTS.md`) - **Effective Price:** $0.056 per image (1MP $0.056 · 2MP $0.08 · 3MP $0.104) ### Tech Stack: Python ### Why It Matters: One key advantage of using FLUX 2 Max (`flux-2-max`) through APIMart is that it provides access to 300+ image, video, and language models via the same base URL and header, eliminating the need to switch models. Additionally, the model is OpenAI-compatible, making it suitable for various applications. The effective price of $0.056 per image is estimated for the entire set, making it a cost-effective solution. The API key can be obtained from the provided link, and the pricing is live and adjustable. The model is designed to be as efficient as possible, with tasks being completed asynchronously by default, allowing for easy integration into workflows without requiring additional setup or complex integrations. ### Repository: Repository: `d79bcrw6n9kd/flux-2-max-prompts-ru` ### Primary Language: Python ### GitHub Stars: 12

### Overview: Nano Banana Pro prompts are tested image prompt recipes designed to create visually appealing and cost-effective renders. This repository provides 8 recipes with examples, aspect ratios, the model ID (`gemini-3-pro-image-preview`), and the cost per image. ### Key Features: - **Model ID:** `gemini-3-pro-image-preview` - **Aspect Ratios:** 1:1, 3:4, 16:9, etc. - **Cost Per Image:** $0.03 (4K $0.04) ### Tech Stack: - Python ### Why It Matters: Nano Banana Pro prompts offer a simple and effective way to create professional-looking images without the need for expensive API keys or monthly fees. The recipes are designed to be easy to follow and maintain, ensuring consistent results across different projects.

### Overview: PangolinGuard is a lightweight, self-hosted AI guardrails model based on ModernBERT, designed to protect against prompt injections and jailbreaks in LLM applications. It closely approximates the performance of Claude 3.7 and Gemini Flash 2.0, maintaining low latency and supporting 8192 input tokens. ### Key Features: - **Lightweight and Self-Hosted**: Runs on a single GPU, making it highly efficient and easy to deploy. - **High Performance**: Close to the performance of Claude 3.7 and Gemini Flash 2.0, ensuring robust security. - **Low Latency**: Operates with low latency, ensuring fast response times. - **Supports 8192 Input Tokens**: Can handle large input sizes, making it suitable for various applications. ### Tech Stack: - **ModernBERT (Large)**: The underlying model architecture. ### Why It Matters: PangolinGuard is crucial for developers and AI practitioners who need to ensure the security and integrity of their AI applications. It helps protect against prompt injections and jailbreaks, ensuring that models remain safe and reliable, especially in critical applications where sensitive information could be at risk.

### Overview: Seedance 1.5 Pro Prompts: 8 Tested Recipes with Outputs <!-- conv-kit:v1 --> <p align="center"> <img src="assets/badges/price.svg" alt="observed unit price"> <img src="assets/badges/billing.svg" alt="billing model"> <img src="assets/badges/compat.svg" alt="OpenAI-compatible endpoint"> </p> <p align="center"> <img src="assets/01-product-ad-ceramic-mug-thumb.jpg" width="820" alt="Sample video frame generated through APIMart"> </p> > **from $0.0104 per second** on the APIMart route (billed by resolution tier) — ein einziger OpenAI-kompatibler Endpoint unter `https://api.apimart.ai/v1`, kein Monatsabo nötig. *(erfasst am 2026-10-12)* **[Get an API key](https://go.apimart.ai/k-42c0f6)** · **[Live pricing](https://go.apimart.ai/k-4ead7f)** · **[Model page](https://go.apimart.ai/k-1ffae8)** · [⚡ 60-second quickstart](#quickstart) **Warum Teams Seedance 1.5 Pro (`seedance-1.5-pro`) über APIMart nutzen** - **Ein Key, der ges

### Overview: SYSTEM-D is a 4-inch touchscreen dashboard designed specifically for vLLM and TensorFold inference servers. It provides real-time monitoring of decode and prefill speeds, KV cache, queues, streams, TTFT, and speculative-decoding acceptance. The dashboard is powered by an ESP32-S3 display board, making it an affordable solution for monitoring your servers without the need for server-side setup. ### Key features: - **Real-time Monitoring:** Displays live metrics from vLLM and TensorFold servers. - **Automatic Detection:** Automatically detects vLLM and TensorFold servers and supports up to 4 servers. - **42 Color Themes:** Offers 42 color themes inspired by btop, including the default SYSTEM-D theme. - **No Server-Side Setup:** The board polls each engine's `/metrics` endpoint over Wi-Fi and calculates metrics independently. - **Honest Numbers:** Displays only measurable data, ensuring accuracy and honesty. ### Tech Stack: - **PlatformIO:** Used for development and configuration. - **ESP32-S3 Display Board:** 480x480 ST7701 IPS panel, 16 MB flash, 8 MB PSRAM. - **Wi-Fi:** Powered by a USB-C data cable connected to a CH340 USB-serial chip. ### Why it Matters: - **Inexpensive and Portable:** Ideal for monitoring servers without the need for additional hardware or expensive software. - **Automated Monitoring:** Automatically detects and monitors vLLM and TensorFold servers, simplifying the setup

Overview: DiFlow is a scalable serving system designed for building and running modular diffusion workflows across GPU clusters. It micro-serves workflows by decomposing them into independently managed model-execution nodes, enabling natural workflow composition, independent model sharing and scaling, and dynamic scheduling based on GPU availability. Key features: - Micro-serves diffusion workflows by decomposing them into independently managed model-execution nodes. - Enables natural workflow composition using Python. - Allows sharing and scaling models independently. - Schedules model executions and adjusts parallelism based on runtime GPU availability. Tech stack: Python Why it matters: DiFlow simplifies the deployment and scaling of complex diffusion workflows, making them more efficient and scalable. By enabling micro-scheduling and independent model management, DiFlow addresses the limitations of traditional monolithic serving systems, providing a more flexible and efficient solution for running large-scale diffusion workflows across GPU clusters.

### Overview: infai is a single repository that provides two terminal tools for local-first AI, enabling seamless inference and workflow management. ### Key features: - **Profiles:** Define launch configurations for llama.cpp and vLLM with context size, GPU layers, KV-cache types, flash attention, sampling, and extra flags. - **Runs dashboard:** Monitor and manage multiple models in a single terminal, with status, uptime, and live tokens per second. - **Live telemetry and logs:** Detailed system and per-model metrics, including CPU, RAM, and VRAM usage, and a scrollable server log. - **Model library:** Automatically links multimodal projectors from model folders. - **Hugging Face downloads:** Search the Hub, download models with resumable, atomic transfers, and keep provenance attached. - **Hardware-fit checks:** Estimate model compatibility with hardware and suggest launch settings. - **11 themes:** Choose from tokyonight, everforest, onedark, rosepine, gruvbox, catppuccin, nord, dracula, kanagawa, solarized, and monokai. ### Tech stack: Go ### Why it matters: infai simplifies local AI inference and workflow management, making it easier to experiment with different models and configurations. ### Repository: dipankardas011/infai ### GitHub stars: 12 ### GitHub description: Local-first AI in your terminal. infai: run llama.cpp and vLLM models from one TUI with launch profiles, live telemetry, and Hugging Face