Open Source AI Models - Page 9
Sort By:
-
Gen AI apps are built with MongoDB AtlasMongoDB Atlas is the developer-friendly database used to build, scale, and run gen AI and LLM-powered apps—without needing a separate vector database. Atlas offers built-in vector search, global availability across 115+ regions, and flexible document modeling. Start building AI apps faster, all in one place.
-
Keep company data safe with Chrome EnterpriseMake AI work your way with Chrome Enterprise. Block unapproved sites and set custom data controls that align with your company's policies.
-
1
Qwen2.5-VL-3B-Instruct
Qwen2.5-VL-3B-Instruct: Multimodal model for chat, vision & video
Qwen2.5-VL-3B-Instruct is a 3.75 billion parameter multimodal model by Qwen, designed to handle complex vision-language tasks in both image and video formats. As part of the Qwen2.5 series, it supports image-text-to-text generation with capabilities like chart reading, object localization, and structured data extraction. The model can serve as an intelligent visual agent capable of interacting with digital interfaces and understanding long-form videos by dynamically sampling resolution and frame rate. It uses a SwiGLU and RMSNorm-enhanced ViT architecture and introduces mRoPE updates for robust temporal and spatial understanding. The model supports flexible image input (file path, URL, base64) and outputs structured responses like bounding boxes or JSON, making it highly versatile in commercial and research settings. It excels in a wide range of benchmarks such as DocVQA, InfoVQA, and AndroidWorld control tasks. -
2
Qwen2.5-VL-7B-Instruct
Multimodal 7B model for image, video, and text understanding tasks
Qwen2.5-VL-7B-Instruct is a multimodal vision-language model developed by the Qwen team, designed to handle text, images, and long videos with high precision. Fine-tuned from Qwen2.5-VL, this 7-billion-parameter model can interpret visual content such as charts, documents, and user interfaces, as well as recognize common objects. It supports complex tasks like visual question answering, localization with bounding boxes, and structured output generation from documents. The model is also capable of video understanding with dynamic frame sampling and temporal reasoning, enabling it to analyze and respond to long-form videos. Built with an enhanced ViT architecture using window attention, SwiGLU, and RMSNorm, it aligns closely with Qwen2.5 LLM standards. The model demonstrates high performance across benchmarks like DocVQA, ChartQA, and MMStar, and even functions as a tool-using visual agent. -
3
Qwen3 Embedding
Designed for text embedding and ranking tasks
Qwen3-Embedding is a model series from the Qwen family designed specifically for text embedding and ranking tasks. It builds upon the Qwen3 base/dense models and offers several sizes (0.6B, 4B, 8B parameters), for both embedding and reranking, with high multilingual capability, long‐context understanding, and reasoning. It achieves state-of-the-art performance on benchmarks like MTEB (Multilingual Text Embedding Benchmark) and supports instruction-aware embedding (i.e. embedding task instructions along with queries) and flexible embedding/vector dimension definitions. It is meant for tasks such as text retrieval, classification, clustering, bitext mining, and code retrieval. -
4
Qwen3-Next
Qwen3-Next: 80B instruct LLM with ultra-long context up to 1M tokens
Qwen3-Next-80B-A3B-Instruct is the flagship release in the Qwen3-Next series, designed as a next-generation foundation model for ultra-long context and efficient reasoning. With 80B total parameters and 3B activated at a time, it leverages hybrid attention (Gated DeltaNet + Gated Attention) and a high-sparsity Mixture-of-Experts architecture to achieve exceptional efficiency. The model natively supports a context length of 262K tokens and can be extended up to 1 million tokens using RoPE scaling (YaRN), making it highly capable for processing large documents and extended conversations. Multi-Token Prediction (MTP) boosts both training and inference, while stability optimizations such as weight-decayed and zero-centered layernorm ensure robustness. Benchmarks show it performs comparably to larger models like Qwen3-235B on reasoning, coding, multilingual, and alignment tasks while requiring only a fraction of the training cost. -
Simple, Secure Domain RegistrationRegister or renew your domain and pay only what we pay. No markups, hidden fees, or surprise add-ons. Choose from over 400 TLDs (.com, .ai, .dev). Every domain is integrated with Cloudflare's industry-leading DNS, CDN, and free SSL to make your site faster and more secure. Simple, secure, at-cost domain registration.
-
5
Ring
Ring is a reasoning MoE LLM provided and open-sourced by InclusionAI
Ring is a reasoning Mixture-of-Experts (MoE) large language model (LLM) developed by inclusionAI. It is built from or derived from Ling. Its design emphasizes reasoning, efficiency, and modular expert activation. In its “flash” variant (Ring-flash-2.0), it optimizes inference by activating only a subset of experts. It applies reinforcement learning/reasoning optimization techniques. Its architectures and training approaches are tuned to enable efficient and capable reasoning performance. Reasoning-optimized model with reinforcement learning enhancements. Efficient architecture and memory design for large-scale reasoning. If you are located in mainland China, we also provide the model on ModelScope.cn to speed up the download process. -
6
SG2Im
Code for "Image Generation from Scene Graphs", Johnson et al, CVPR 201
sg2im is a research codebase that learns to synthesize images from scene graphs—structured descriptions of objects and their relationships. Instead of conditioning on free-form text alone, it leverages graph structure to control layout and interactions, generating scenes that respect constraints like “person left of dog” or “cup on table.” The pipeline typically predicts object layouts (bounding boxes and masks) from the graph, then renders a realistic image conditioned on those layouts. This separation lets the model reason about geometry and composition before committing to texture and color, improving spatial fidelity. The repository includes training code, datasets, and evaluation scripts so researchers can reproduce baselines and extend components such as the graph encoder or image generator. In practice, sg2im demonstrates how structured semantics can guide generative models to produce controllable, compositional imagery. -
7
Seamless Communication
Foundational Models for State-of-the-Art Speech and Text Translation
Seamless Communication is a research project focused on building more integrated, low-latency multimodal communication between humans and AI agents. The motivation is to move beyond “text in, text out” and enable direct, live, multi-turn exchange involving language, gesture, gaze, vision, and modality switching without user friction. The system architecture includes a real-time multimodal signal pipeline for audio, video, and sensor data, a dialog manager that can decide when to act (speak, gesture, point) or query, and a cross-modal reasoning layer that fuses perception with semantic context. The research prototype includes components for visual grounding (understanding when a user references something in view), gesture recognition and synthesis, and turn-taking mechanisms that mirror human conversational timing. Because latency and synchronization are critical, the codebase invests in asynchronous scheduling, overlap of perception and reasoning, and fast fallback responses. -
8
Stable Virtual Camera
Stable Virtual Camera: Generative View Synthesis with Diffusion Models
Stable Virtual Camera is a multi-view diffusion model developed by Stability AI that transforms 2D images into immersive 3D videos with realistic depth and perspective. Unlike traditional methods that require complex reconstruction or scene-specific optimization, this model allows users to generate novel views from any number of input images and define custom camera trajectories, enabling dynamic exploration of scenes. It supports various aspect ratios and can produce 3D-consistent videos up to 1,000 frames, making it a versatile tool for creators seeking to enhance visual storytelling. -
9
Stable-Dreamfusion
Text-to-3D & Image-to-3D & Mesh Exportation with NeRF + Diffusion
A pytorch implementation of the text-to-3D model Dreamfusion, powered by the Stable Diffusion text-to-2D model. This project is a work-in-progress and contains lots of differences from the paper. The current generation quality cannot match the results from the original paper, and many prompts still fail badly! Since the Imagen model is not publicly available, we use Stable Diffusion to replace it (implementation from diffusers). Different from Imagen, Stable-Diffusion is a latent diffusion model, which diffuses in a latent space instead of the original image space. Therefore, we need the loss to propagate back from the VAE's encoder part too, which introduces extra time costs in training. We use the multi-resolution grid encoder to implement the NeRF backbone (implementation from torch-ngp), which enables much faster rendering. -
Level Up Your Cyber Defense with External Threat ManagementMove beyond alerts. Gain full visibility, context, and control over your external attack surface to stay ahead of every threat.
-
10
StarSpace
Learning embeddings for classification, retrieval and ranking
StarSpace is a general-purpose embedding-based learning framework that trains embeddings for entities (words, sentences, users, items) under various supervision signals (classification, ranking, matching). Instead of focusing on one task, StarSpace supports multi-task and multi-domain setups—for instance, you can train embeddings so that textual queries match item descriptions, sentences map to labels, or users align with liked items in the same embedding space. The training objective is contrastive: for a given query embedding, positive and negative examples are sampled and the model is optimized to score positive higher than negatives. The library supports a variety of tasks (text classification, nearest-neighbor search, recommendation, entity linking) with simple configuration. It includes efficient batching, negative sampling strategies, and on-the-fly embedding updates. -
11
Style Aligned
Official code for Style Aligned Image Generation via Shared Attention
StyleAligned is a diffusion-model editing technique and codebase that preserves the visual “style” of an original image while applying new semantic edits driven by text. Instead of fully re-generating an image—and risking changes to lighting, texture, or rendering choices—the method aligns internal features across denoising steps so the target edit inherits the source style. This alignment acts like a constraint on the model’s evolution, steering composition, palette, and brushwork even as objects or attributes change. The result is more consistent edits across a set, which is crucial for workflows like product variations, character sheets, or brand-coherent art. The repository provides reproducible scripts, reference prompts, and guidance for tuning strengths so users can dial in subtle retouches or bolder substitutions. Because it builds on widely used diffusion checkpoints, creators can integrate it without training or dataset collection. -
12
Surya
Implementation of the Surya Foundation Model for Heliophysics
Surya is an open‑source, AI‑based foundation model for heliophysics developed collaboratively by NASA (via the IMPACT AI team) and IBM. Named after the Sanskrit word for “sun,” Surya is trained on nine years of high‑resolution solar imagery from NASA’s Solar Dynamics Observatory (SDO). It is designed to forecast solar phenomena—such as flares, solar wind, irradiance, and active region behavior—by predicting future solar images with a sophisticated long–short vision transformer architecture, thereby enabling improved space weather forecasting. Foresees solar flares, wind, EUV spectra, and active region formation in advance. Achieves approximately 16% improvement in forecasting accuracy over traditional methods. 366-million‑parameter foundation model capturing general-purpose solar representations. -
13
Tencent-Hunyuan-Large
Open-source large language model family from Tencent Hunyuan
Tencent-Hunyuan-Large is the flagship open-source large language model family from Tencent Hunyuan, offering both pre-trained and instruct (fine-tuned) variants. It is designed with long-context capabilities, quantization support, and high performance on benchmarks across general reasoning, mathematics, language understanding, and Chinese / multilingual tasks. It aims to provide competitive capability with efficient deployment and inference. FP8 quantization support to reduce memory usage (~50%) while maintaining precision. High benchmarking performance on tasks like MMLU, MATH, CMMLU, C-Eval, etc. -
14
Tiktoken
tiktoken is a fast BPE tokeniser for use with OpenAI's models
tiktoken is a high-performance, tokenizer library (based on byte-pair encoding, BPE) designed for use with OpenAI’s models. It handles encoding and decoding text to token IDs efficiently, with minimal overhead. Because tokenization is a fundamental step in preparing text for models, tiktoken is optimized for speed, memory, and correctness in model contexts (e.g. matching OpenAI’s internal tokenization). The repo supports multiple encodings (e.g. “cl100k_base”) and lets users switch encoding names to match different model contexts. It also offers extension mechanisms so that custom encodings can be registered. Internally, it includes the core tokenizer logic (often implemented in Rust or efficient lower-level code), APIs for encoding, decoding, and counting tokens, and binding layers to Python (and sometimes other languages) for easy use. -
15
TimeSformer
The official pytorch implementation of our paper
TimeSformer is a vision transformer architecture for video that extends the standard attention mechanism into spatiotemporal attention. The model alternates attention along spatial and temporal dimensions (or designs variants like divided attention) so that it can capture both appearance and motion cues in video. Because the attention is global across frames, TimeSformer can reason about dependencies across long time spans, not just local neighborhoods. The official implementation in PyTorch provides configurations, pretrained models, and training scripts that make it straightforward to evaluate or fine-tune on video datasets. TimeSformer was influential in showing that pure transformer architectures—without convolutional backbones—can perform strongly on video classification tasks. Its flexible attention design allows experimenting with different factoring (spatial-then-temporal, joint, etc.) to trade off compute, memory, and accuracy. -
16
ToMe (Token Merging)
A method to increase the speed and lower the memory footprint
ToMe (Token Merging) is a PyTorch-based optimization framework designed to significantly accelerate Vision Transformer (ViT) architectures without retraining. Developed by researchers at Facebook (Meta AI), ToMe introduces an efficient technique that merges similar tokens within transformer layers, reducing redundant computation while preserving model accuracy. This approach differs from token pruning, which removes background tokens entirely; instead, ToMe merges tokens based on feature similarity, allowing it to compress both foreground and background information efficiently. ToMe integrates seamlessly into existing transformer models such as DeiT, MAE, SWAG, and timm ViTs, offering 2–3x speedups during inference and substantial efficiency gains during training. The method can be applied dynamically at inference time or incorporated into training for improved performance. -
17
UCO3D
Uncommon Objects in 3D dataset
uCO3D is a large-scale 3D vision dataset and toolkit centered on turn-table videos of everyday objects drawn from the LVIS taxonomy. It provides about 170,000 full videos per object instance rather than still frames, along with per-video annotations including object masks, calibrated camera poses, and multiple flavors of point clouds. Each sequence also ships with a precomputed 3D Gaussian Splat reconstruction, enabling fast, differentiable rendering workflows and modern implicit/point-based modeling experiments. The repository includes automated downloaders with checksum verification, fine-grained controls to fetch only selected modalities or super-categories, and a lightweight Python API for loading frames, geometry, and splats on demand. Metadata is indexed in SQLite for quick queries at scale, and helper builders handle alignment, undistortion, frame extraction from videos, and cropping around the object. -
18
Universal Sentence Encoder
Encoder of greater-than-word length text trained on a variety of data
The Universal Sentence Encoder (USE) is a pre-trained deep learning model designed to encode sentences into fixed-length embeddings for use in various natural language processing (NLP) tasks. It leverages Transformer and Deep Averaging Network (DAN) architectures to generate embeddings that capture the semantic meaning of sentences. The model is designed for tasks like sentiment analysis, semantic textual similarity, and clustering, and provides high-quality sentence representations in a computationally efficient manner. -
19
VaultGemma
VaultGemma: 1B DP-trained Gemma variant for private NLP tasks
VaultGemma is a sub-1B parameter variant of Google’s Gemma family that is pre-trained from scratch with Differential Privacy (DP), providing mathematically backed guarantees that its outputs do not reveal information about any single training example. Using DP-SGD with a privacy budget across a large English-language corpus (web documents, code, mathematics), it prioritizes privacy over raw utility. The model follows a Gemma-2–style architecture, outputs text from up to 1,024 input tokens, and is intended to be instruction-tuned for downstream language understanding and generation tasks. Training ran on TPU v6e using JAX and Pathways with privacy-preserving algorithms (DP-SGD, truncated Poisson subsampling) and DP scaling laws to balance compute and privacy budgets. Benchmarks on the 1B pre-trained checkpoint show expected utility trade-offs (e.g., HellaSwag 10-shot 39.09, BoolQ 0-shot 62.04, PIQA 0-shot 68.00), reflecting its privacy-first design. -
20
VibeVoice
VibeVoice: Open-source multi-speaker long-form text-to-speech model
VibeVoice-1.5B is Microsoft’s frontier open-source text-to-speech (TTS) model designed for generating expressive, long-form, multi-speaker conversational audio such as podcasts. Unlike traditional TTS systems, it excels in scalability, speaker consistency, and natural turn-taking for up to 90 minutes of continuous speech with as many as four distinct speakers. A key innovation is its use of continuous acoustic and semantic speech tokenizers operating at an ultra-low frame rate of 7.5 Hz, enabling high audio fidelity with efficient processing of long sequences. The model integrates a Qwen2.5-based large language model with a diffusion head to produce realistic acoustic details and capture conversational context. Training involved curriculum learning with increasing sequence lengths up to 65K tokens, allowing VibeVoice to handle very long dialogues effectively. Safety mechanisms include an audible disclaimer and imperceptible watermarking in all generated audio to mitigate misuse risks. -
21
Video Pre-Training
Learning to Act by Watching Unlabeled Online Videos
The Video PreTraining (VPT) repository provides code and model artifacts for a project where agents learn to act by watching human gameplay videos—specifically, gameplay of Minecraft—using behavioral cloning. The idea is to learn general priors of control from large-scale, unlabeled video data, and then optionally fine-tune those priors for more goal-directed behavior via environment interaction. The repository contains demonstration models of different widths, fine-tuned variants (e.g. for building houses or early-game tasks), and inference scripts that instantiate agents from pretrained weights. Key modules include the behavioral cloning logic, the agent wrapper, and data loading pipelines (with an accessible skeleton for loading Minecraft demonstration data). The repo also includes a run_agent.py script for testing an agent interactively, and an agent.py module encapsulating the control logic. -
22
VisualGLM-6B
Chinese and English multimodal conversational language model
VisualGLM-6B is an open-source multimodal conversational language model developed by ZhipuAI that supports both images and text in Chinese and English. It builds on the ChatGLM-6B backbone, with 6.2 billion language parameters, and incorporates a BLIP2-Qformer visual module to connect vision and language. In total, the model has 7.8 billion parameters. Trained on a large bilingual dataset — including 30 million high-quality Chinese image-text pairs from CogView and 300 million English pairs — VisualGLM-6B is designed for image understanding, description, and question answering. Fine-tuning on long visual QA datasets further aligns the model’s responses with human preferences. The repository provides inference APIs, command-line demos, web demos, and efficient fine-tuning options like LoRA, QLoRA, and P-tuning. It also supports quantization down to INT4, enabling local deployment on consumer GPUs with as little as 6.3 GB VRAM. -
23
Watermark Anything
Official implementation of Watermark Anything with Localized Messages
Watermark Anything (WAM) is an advanced deep learning framework for embedding and detecting localized watermarks in digital images. Developed by Facebook Research, it provides a robust, flexible system that allows users to insert one or multiple watermarks within selected image regions while maintaining visual quality and recoverability. Unlike traditional watermarking methods that rely on uniform embedding, WAM supports spatially localized watermarks, enabling targeted protection of specific image regions or objects. The model is trained to balance imperceptibility, ensuring minimal visual distortion, with robustness against transformations and edits such as cropping or motion. -
24
bart-large-cnn
Summarization model fine-tuned on CNN/DailyMail articles
facebook/bart-large-cnn is a large-scale sequence-to-sequence transformer model developed by Meta AI and fine-tuned specifically for abstractive text summarization. It uses the BART architecture, which combines a bidirectional encoder (like BERT) with an autoregressive decoder (like GPT). Pre-trained on corrupted text reconstruction, the model was further trained on the CNN/DailyMail dataset—a collection of news articles paired with human-written summaries. It performs particularly well in generating concise, coherent, and human-readable summaries from longer texts. Its architecture allows it to model both language understanding and generation tasks effectively. The model supports usage in PyTorch, TensorFlow, and JAX, and is integrated with the Hugging Face pipeline API for simple deployment. Due to its size and performance, it's widely used in real-world summarization applications such as news aggregation, legal document condensing, and content creation. -
25
bge-base-en-v1.5
Efficient English embedding model for semantic search and retrieval
bge-base-en-v1.5 is an English sentence embedding model from BAAI optimized for dense retrieval tasks, part of the BGE (BAAI General Embedding) family. It is a fine-tuned BERT-based model designed to produce high-quality, semantically meaningful embeddings for tasks like semantic similarity, information retrieval, classification, and clustering. This version (v1.5) improves retrieval performance and stabilizes similarity score distribution without requiring instruction-based prompts. With 768 embedding dimensions and a maximum sequence length of 512 tokens, it achieves strong performance across multiple MTEB benchmarks, nearly matching larger models while maintaining efficiency. It supports use via SentenceTransformers, Hugging Face Transformers, FlagEmbedding, and ONNX for various deployment scenarios. Typical usage includes normalizing output embeddings and calculating cosine similarity via dot product for ranking.