TensorRT LLM is an inference performance framework for large language models (LLMs) on NVIDIA GPUs, developed by NVIDIA. Based on the PyTorch architecture, TensorRT LLM provides an efficient and easy-to-use Python API...
Qwen3-Max-Thinking is Alibaba's latest flagship inference model for the Thousand Questions platform, boasting over 1 trillion parameters and 36TB of pre-trained tokens. Through test-time scaling and reinforcement learning, the model significantly improves inference performance and efficiency...
HunyuanImage 3.0-Instruct is an advanced image generation and editing model launched by Tencent. Based on an 80-parameter Hybrid Expert (MoE) architecture, the model can deeply understand user-input images...
SGLang is an open-source, high-performance inference framework for large-scale language and multimodal models. The framework provides low-latency, high-throughput inference services and supports deployments from single GPUs to large-scale distributed clusters.
vLLM is an open-source Large Language Model (LLM) inference and deployment framework developed by the Sky Computing Lab at UC Berkeley. It supports providing users with fast, low-cost model services. vLLM utilizes innovative memory management techniques...
Clawdbot is an open-source personal AI assistant that runs on a local device and interacts with users through chat applications such as WhatsApp and Telegram.
OpenJudge is an open-source AI application evaluation framework that solves the transition challenge from prototype to production environments. Through a systematic evaluation mechanism, the framework helps developers quantitatively assess the performance of their AI applications, ensuring their effectiveness in complex business scenarios...
D4RT (Dynamic 4D Reconstruction and Tracking) is a dynamic 4D reconstruction model developed by Google DeepMind. The model integrates 3D reconstruction, camera tracking, and dynamic object tracking through a unified spatiotemporal query interface...
FlowAct-R1 is a real-time interactive digital human video generation framework launched by ByteDance. It only requires a single reference image and audio to support the streaming generation of full-body dynamic videos of unlimited length.
Agentation is an open-source AI programming collaboration tool that solves the problem of inefficient communication between developers and AI programming assistants. The tool allows users to directly annotate issues on a webpage, translating visual feedback into readable code...
Qwen3-TTS is a series of open-source speech generation models from Qwen, possessing powerful capabilities in timbre cloning, creation, and voice control. The model is based on the innovative Qwen3-TTS-Tokenizer-12Hz multi-codebook speech encoder, achieving efficient speech...
Being-H0.5 is a general-purpose robot model developed by Lu Zongqing's team. It can solve the problem of strategy transfer between different hardware forms of robots by using human prior knowledge and unified action alignment.
LightOnOCR-2-1B is a high-efficiency OCR model launched by LightOnAI. With only 1B of parameters, it performs exceptionally well in complex document processing and excels at recognizing academic papers, mathematical formulas, and complex tables.
AgentCPM-Report is a localized deep research agent jointly developed by the Natural Language Processing Laboratory of Tsinghua University, Renmin University of China, Wallfacer Intelligence, and the OpenBMB open-source community. It is based on the MiniCPM4.1 model with 8B parameters.
VibeVoice-ASR is an advanced speech recognition model open-sourced by Microsoft, designed specifically for processing long audio clips up to 60 minutes long. The model can process the entire audio segment at once, preserving the global context and avoiding the limitations of segmented processing common in traditional models...
Model1 is a mysterious model that DeepSeek has updated and revealed in the FlashMLA codebase. It may be the internal codename or the first engineering version of the next-generation flagship model, DeepSeek-V4.
Chroma 1.0 is FlashLabs' first open-source real-time end-to-end speech dialogue model, featuring low-latency interaction, high-fidelity personalized voice cloning, and strong conversational capabilities. The model tightly couples speech understanding and generation, employing a 1:2 text-to-speech ratio...
Step3-VL-10B is an open-source multimodal model with only 10 parameters, developed by Step3 StarCraft. It achieves the performance level of a 200-parameter model in tasks such as visual perception, logical reasoning, mathematical competitions, and general dialogue.
EmbodiChain is a cross-dimensional, open-source embodied intelligence learning platform that drives the development of embodied intelligence through generative simulation data. It automatically creates physically accurate 3D scenes and tasks, combining online data streams and self-healing...
json-render is an open-source project by Vercel that solves the problem of uncontrollable AI-generated UIs. json-render defines a Catalog to constrain AI to only generate JSON data conforming to a specific schema, using...
COTA is a new type of game intelligent agent launched by Hyperparameter Technology. Driven by a Large Language Model (LLM), it possesses cognitive, operational, tactical, and auxiliary capabilities. COTA breaks through traditional reinforcement learning and supervised learning models, achieving its goals through innovative architecture...
x-Algorithm is Elon Musk's open-source recommendation algorithm for the x platform, the core system for providing personalized content for the "For You" news feed. It combines posts from accounts followed by the user (In-Network) and global content retrieved through machine learning (Out-Network)...
NVIDIA PersonaPlex is a full-duplex dialogue AI model from NVIDIA, capable of simultaneously listening and speaking, and handling interruptions, pauses, and responses in natural conversations. Users can customize the persona and voice using voice and text prompts...
GLM-4.7-Flash is a free, open-source hybrid thinking model from Zhipu, featuring 30 billion parameters and 3 billion activation parameters. The model achieves a balance between performance and efficiency, performing exceptionally well in various scenarios, including programming, Chinese writing, and translation.