Open Source AI Models - Page 5
Sort By:
-
Photo and Video Editing APIs and SDKsUnlock Picsart's full editing suite by embedding our Editor SDK directly into your platform. Offer your users the power of a full design suite without leaving your site.
-
Build Securely on Azure with Proven FrameworksMoving to the cloud brings new challenges. How can you manage a larger attack surface while ensuring great network performance? Turn to Fortinet’s Tested Reference Architectures, blueprints for designing and securing cloud environments built by cybersecurity experts. Learn more and explore use cases in this white paper.
-
1
Apple Neural Engine (ANE) Transformers
Reference implementation of the Transformer architecture optimized
ANE Transformers is a reference PyTorch implementation of Transformer components optimized for Apple Neural Engine on devices with A14 or newer and on Macs with M1 or newer chips. It demonstrates how to structure attention and related layers to achieve substantial speedups and lower peak memory compared to baseline implementations when deployed to ANE. The repository targets practitioners who want to keep familiar PyTorch modeling while preparing models for Core ML/ANE execution paths. Documentation highlights reported improvements in throughput and memory residency, while releases track incremental fixes and packaging updates. The project sits alongside related Apple ML repos that focus on deploying attention-based models efficiently to ANE-equipped hardware. In short, it’s a practical blueprint for adapting Transformers to Apple’s dedicated ML accelerator without rewriting entire model stacks. -
2
BLEURT-20-D12
Custom BLEURT model for evaluating text similarity using PyTorch
BLEURT-20-D12 is a PyTorch implementation of BLEURT, a model designed to assess the semantic similarity between two text sequences. It serves as an automatic evaluation metric for natural language generation tasks like summarization and translation. The model predicts a score indicating how similar a candidate sentence is to a reference sentence, with higher scores indicating greater semantic overlap. Unlike standard BLEURT models from TensorFlow, this version is built from a custom PyTorch transformer library. It requires installing the model-specific library from GitHub to function properly. Once set up, it can be used to compute similarity scores with minimal code. BLEURT-20-D12 enables more flexible deployment in PyTorch-based workflows for evaluating language generation outputs. -
3
BioEmu
Inference code for scalable emulation of protein equilibrium ensembles
Biomolecular Emulator (BioEmu for short) is a model that samples from the approximated equilibrium distribution of structures for a protein monomer, given its amino acid sequence. By default, unphysical structures (steric clashes or chain discontinuities) will be filtered out, so you will typically get fewer samples in the output than requested. The difference can be very large if your protein has large disordered regions, which are very likely to produce clashes. BioEmu outputs structures in backbone frame representation. To reconstruct the side-chains, several tools are available. As an example, we interface with HPacker to conduct side-chain reconstruction and also provide basic tooling for running a short molecular dynamics (MD) equilibration. -
4
Bio_ClinicalBERT
ClinicalBERT model trained on MIMIC notes for clinical NLP tasks
Bio_ClinicalBERT is a domain-specific language model tailored for clinical natural language processing (NLP), extending BioBERT with additional training on clinical notes. It was initialized from BioBERT-Base v1.0 and further pre-trained on all clinical notes from the MIMIC-III database (~880M words), which includes ICU patient records. The training focused on improving performance in tasks like named entity recognition and natural language inference within the healthcare domain. Notes were processed using rule-based sectioning and tokenized with SciSpacy. Training was done for 150,000 steps using a batch size of 32, max sequence length of 128, and a masked language modeling objective with a 0.15 mask probability. Bio_ClinicalBERT is available through Hugging Face's Transformers library for easy integration. It supports medical AI research and applications involving electronic health record understanding, clinical decision support, and biomedical information extraction. -
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.
-
5
CLIP
CLIP, Predict the most relevant text snippet given an image
CLIP (Contrastive Language-Image Pretraining) is a neural model that links images and text in a shared embedding space, allowing zero-shot image classification, similarity search, and multimodal alignment. It was trained on large sets of (image, caption) pairs using a contrastive objective: images and their matching text are pulled together in embedding space, while mismatches are pushed apart. Once trained, you can give it any text labels and ask it to pick which label best matches a given image—even without explicit training for that classification task. The repository provides code for model architecture, preprocessing transforms, evaluation pipelines, and example inference scripts. Because it generalizes to arbitrary labels via text prompts, CLIP is a powerful tool for tasks that involve interpreting images in terms of descriptive language. -
6
CLIP-ViT-bigG-14-laion2B-39B-b160k
CLIP ViT-bigG/14: Zero-shot image-text model trained on LAION-2B
CLIP-ViT-bigG-14-laion2B-39B-b160k is a powerful vision-language model trained on the English subset of the LAION-5B dataset using the OpenCLIP framework. Developed by LAION and trained by Mitchell Wortsman on Stability AI’s compute infrastructure, it pairs a ViT-bigG/14 vision transformer with a text encoder to perform contrastive learning on image-text pairs. This model excels at zero-shot image classification, image-to-text and text-to-image retrieval, and can be adapted for tasks such as image captioning or generation guidance. It achieves an impressive 80.1% top-1 accuracy on ImageNet-1k without any fine-tuning, showcasing its robustness in open-domain settings. Its training dataset is uncurated and web-sourced, meaning it reflects the biases and risks of large-scale internet data. The model is intended for research use and is not recommended for real-world deployment without domain-specific testing and safety evaluations. -
7
ChatGLM2-6B
An Open Bilingual Chat LLM | Open Source Bilingual Conversation LLM
ChatGLM2-6B is an advanced open-source bilingual dialogue model developed by THUDM. It is the second iteration of the ChatGLM series, designed to offer enhanced performance while maintaining the strengths of its predecessor, including smooth conversation flow and low deployment barriers. The model is fine-tuned for both Chinese and English languages, making it a versatile tool for various multilingual applications. ChatGLM2-6B aims to push the boundaries of natural language understanding and generation, offering improved accuracy and user experience compared to earlier models. -
8
ChatGPT Retrieval Plugin
The ChatGPT Retrieval Plugin lets you easily find personal documents
The chatgpt-retrieval-plugin repository implements a semantic retrieval backend that lets ChatGPT (or GPT-powered tools) access private or organizational documents in natural language by combining vector search, embedding models, and plugin infrastructure. It can serve as a custom GPT plugin or function-calling backend so that a chat session can “look up” relevant documents based on user queries, inject those results into context, and respond more knowledgeably about a private knowledge base. The repo provides code for ingestion pipelines (embedding documents), APIs for querying, local server components, and privacy / PII detection modules. It also contains plugin manifest files (OpenAPI spec, plugin JSON) so that the retrieval backend can be registered in a plugin ecosystem. Because retrieval is often needed to make LLMs “know what’s in your docs” without leaking everything, this plugin aims to be a secure, flexible building block for retrieval-augmented generation (RAG) systems. -
9
Chinese-LLaMA-Alpaca 2
Chinese LLaMA-2 & Alpaca-2 Large Model Phase II Project
This project is developed based on the commercially available large model Llama-2 released by Meta. It is the second phase of the Chinese LLaMA&Alpaca large model project. The Chinese LLaMA-2 base model and the Alpaca-2 instruction fine-tuning large model are open-sourced. These models expand and optimize the Chinese vocabulary on the basis of the original Llama-2, use large-scale Chinese data for incremental pre-training, and further improve the basic semantics and command understanding of Chinese. Performance improvements. The related model supports FlashAttention-2 training, supports 4K context and can be extended up to 18K+ through the NTK method. -
Build Securely on AWS with Proven FrameworksMoving to the cloud brings new challenges. How can you manage a larger attack surface while ensuring great network performance? Turn to Fortinet’s Tested Reference Architectures, blueprints for designing and securing cloud environments built by cybersecurity experts. Learn more and explore use cases in this white paper.
-
10
Clay Foundation Model
The Clay Foundation Model - An open source AI model and interface
The Clay Foundation Model is an open-source AI model and interface designed to provide comprehensive data and insights about Earth. It aims to serve as a foundational tool for environmental monitoring, research, and decision-making by integrating various data sources and offering an accessible platform for analysis. -
11
CodeGeeX
CodeGeeX: An Open Multilingual Code Generation Model (KDD 2023)
CodeGeeX is a large-scale multilingual code generation model with 13 billion parameters, trained on 850B tokens across more than 20 programming languages. Developed with MindSpore and later made PyTorch-compatible, it is capable of multilingual code generation, cross-lingual code translation, code completion, summarization, and explanation. It has been benchmarked on HumanEval-X, a multilingual program synthesis benchmark introduced alongside the model, and achieves state-of-the-art performance compared to other open models like InCoder and CodeGen. CodeGeeX also powers IDE plugins for VS Code and JetBrains, offering features like code completion, translation, debugging, and annotation. The model supports Ascend 910 and NVIDIA GPUs, with optimizations like quantization and FasterTransformer acceleration for faster inference. -
12
CodeGeeX2
CodeGeeX2: A More Powerful Multilingual Code Generation Model
CodeGeeX2 is the second-generation multilingual code generation model from ZhipuAI, built upon the ChatGLM2-6B architecture and trained on 600B code tokens. Compared to the first generation, it delivers a significant boost in programming ability across multiple languages, outperforming even larger models like StarCoder-15B in some benchmarks despite having only 6B parameters. The model excels at code generation, translation, summarization, debugging, and comment generation, and it supports over 100 programming languages. With improved inference efficiency, quantization options, and multi-query/flash attention, CodeGeeX2 achieves faster generation speeds and lightweight deployment, requiring as little as 6GB GPU memory at INT4 precision. Its backend powers the CodeGeeX IDE plugins for VS Code, JetBrains, and other editors, offering developers interactive AI assistance with features like infilling and cross-file completion. -
13
CogVLM2
GPT4V-level open-source multi-modal model based on Llama3-8B
CogVLM2 is the second generation of the CogVLM vision-language model series, developed by ZhipuAI and released in 2024. Built on Meta-Llama-3-8B-Instruct, CogVLM2 significantly improves over its predecessor by providing stronger performance across multimodal benchmarks such as TextVQA, DocVQA, and ChartQA, while introducing extended context length support of up to 8K tokens and high-resolution image input up to 1344×1344. The series includes models for both image understanding and video understanding, with CogVLM2-Video supporting up to 1-minute videos by analyzing keyframes. It supports bilingual interaction (Chinese and English) and has open-source versions optimized for dialogue and video comprehension. Notably, the Int4 quantized version allows efficient inference on GPUs with only 16GB of memory. The repository offers demos, API servers, fine-tuning examples, and integration with OpenAI API-compatible endpoints, making it accessible for both researchers and developers. -
14
Consistency Models
Official repo for consistency models
consistency_models is the repository for Consistency Models, a new family of generative models introduced by OpenAI that aim to generate high-quality samples by mapping noise directly into data — circumventing the need for lengthy diffusion chains. It builds on and extends diffusion model frameworks (e.g. based on the guided-diffusion codebase), adding techniques like consistency distillation and consistency training to enable fast, often one-step, sample generation. The repo is implemented in PyTorch and includes support for large-scale experiments on datasets like ImageNet-64 and LSUN variants. It also contains checkpointed models, evaluation scripts, and variants of sampling / editing algorithms described in the paper. Because consistency models reduce the number of inference steps, they are promising for real-time or low-latency generative systems. -
15
ConvNeXt V2
Code release for ConvNeXt V2 model
ConvNeXt V2 is an evolution of the ConvNeXt architecture that co-designs convolutional networks alongside self-supervised learning. The V2 version introduces a fully convolutional masked autoencoder (FCMAE) framework where parts of the image are masked and the network reconstructs the missing content, marrying convolutional inductive bias with powerful pretraining. A key innovation is a new Global Response Normalization (GRN) layer added to the ConvNeXt backbone, which enhances feature competition across channels. The result is a convnet that competes strongly with transformer architectures on recognition benchmarks while being efficient and hardware-friendly. The repository provides official PyTorch implementations for multiple model sizes (Atto, Femto, Pico, up through Huge), conversion from JAX weights, code for pretraining/fine-tuning, and pretrained checkpoints. It supports both self-supervised pretraining and supervised fine-tuning. -
16
DeepGEMM
Clean and efficient FP8 GEMM kernels with fine-grained scaling
DeepGEMM is a specialized CUDA library for efficient, high-performance general matrix multiplication (GEMM) operations, with particular focus on low-precision formats such as FP8 (and experimental support for BF16). The library is designed to work cleanly and simply, avoiding overly templated or heavily abstracted code, while still delivering performance that rivals expert-tuned libraries. It supports both standard and “grouped” GEMMs, which is useful for architectures like Mixture of Experts (MoE) that require segmented matrix multiplications. One distinguishing aspect is that DeepGEMM compiles its kernels at runtime (via a lightweight Just-In-Time (JIT) module), so users don’t need to precompile CUDA kernels before installation. Despite its lean design, it includes scaling strategies (fine-grained scaling) and optimizations inspired by cutting edge systems (drawing from ideas in CUTLASS, CuTe) but in a more streamlined form. -
17
DeepSeek LLM
DeepSeek LLM: Let there be answers
The DeepSeek-LLM repository hosts the code, model files, evaluations, and documentation for DeepSeek’s LLM series (notably the 67B Chat variant). Its tagline is “Let there be answers.” The repo includes an “evaluation” folder (with results like math benchmark scores) and code artifacts (e.g. pre-commit config) that support model development and deployment. According to the evaluation files, DeepSeek LLM 67B Chat achieves strong performance on math benchmarks under both chain-of-thought (CoT) and tool-assisted reasoning modes. The model is trained from scratch, reportedly on a vast multilingual + code + reasoning dataset, and competes with other open or open-weight models. The architecture mirrors established decoder-only transformer families: pre-norm structure, rotational embeddings (RoPE), grouped query attention (GQA), and mixing in languages and tasks. It supports both “Base” (foundation model) and “Chat” (instruction / conversation tuned) variants. -
18
DeepSeek MoE
Towards Ultimate Expert Specialization in Mixture-of-Experts Language
DeepSeek-MoE (“DeepSeek MoE”) is the DeepSeek open implementation of a Mixture-of-Experts (MoE) model architecture meant to increase parameter efficiency by activating only a subset of “expert” submodules per input. The repository introduces fine-grained expert segmentation and shared expert isolation to improve specialization while controlling compute cost. For example, their MoE variant with 16.4B parameters claims comparable or better performance to standard dense models like DeepSeek 7B or LLaMA2 7B using about 40% of the total compute. The repo publishes both Base and Chat variants of the 16B MoE model (deepseek-moe-16b) and provides evaluation results across benchmarks. It also includes a quick start with inference instructions (using Hugging Face Transformers) and guidance on fine-tuning (DeepSpeed, hyperparameters, quantization). The licensing is MIT for code, with a “Model License” applied to the models. -
19
DeepSeek Prover V2
Advancing Formal Mathematical Reasoning via Reinforcement Learning
DeepSeek-Prover-V2 is DeepSeek’s specialized model for formal theorem proving, particularly targeting proof in Lean 4. The repository describes how they use recursive proof decomposition by prompting DeepSeek-V3 to break complex theorems into subgoals, synthesize proof sketches, and then combine them to bootstrap training data. They then fine-tune via reinforcement learning with binary correct/incorrect feedback to integrate informal reasoning with formal proof behavior. The repo releases two model sizes (7B and 671B) and provides evaluation performance (e.g. pass rates on MiniF2F, results on ProverBench) as well as prompt / usage examples for proof generation in Lean 4. It also includes a PDF of the paper or project overview and sample formalization datasets. Because theorem proving is a cutting-edge area in LLM research, Prover-V2 is positioned as a pushing-forward effort in formal reasoning for LLMs. -
20
DeepSeek VL
Towards Real-World Vision-Language Understanding
DeepSeek-VL is DeepSeek’s initial vision-language model that anchors their multimodal stack. It enables understanding and generation across visual and textual modalities—meaning it can process an image + a prompt, answer questions about images, caption, classify, or reason about visuals in context. The model is likely used internally as the visual encoder backbone for agent use cases, to ground perception in downstream tasks (e.g. answering questions about a screenshot). The repository includes model weights (or pointers to them), evaluation metrics on standard vision + language benchmarks, and configuration or architecture files. It also supports inference tools for forwarding image + prompt through the model to produce text output. DeepSeek-VL is a predecessor to their newer VL2 model, and presumably shares core design philosophy but with earlier scaling, fewer enhancements, or capability tradeoffs. -
21
DeepSeek VL2
Mixture-of-Experts Vision-Language Models for Advanced Multimodal
DeepSeek-VL2 is DeepSeek’s vision + language multimodal model—essentially the next-gen successor to their first vision-language models. It combines image and text inputs into a unified embedding / reasoning space so that you can query with text and image jointly (e.g. “What’s going on in this scene?” or “Generate a caption appropriate to context”). The model supports both image understanding (vision tasks) and multimodal reasoning, and is likely used as a component in agent systems to process visual inputs as context for downstream tasks. The repository includes evaluation results (e.g. image/text alignment scores, common VL benchmarks), configuration files, and model weights (where permitted). While the internal architecture details are not fully documented publicly, the repo suggests that VL2 introduces enhancements over prior vision-language models (e.g. better scaling, cross-modal attention, more robust alignment) to improve grounding and multimodal understanding. -
22
DeepSeek-V3.1-Terminus
685B model with improved agents and consistency
DeepSeek-V3.1-Terminus is an updated release in the DeepSeek-V3.1 series, maintaining the original model’s large-scale reasoning and generative capabilities while addressing several key user-reported issues. It improves language consistency, reducing mixed Chinese-English outputs and eliminating abnormal characters, enhancing reliability in multilingual scenarios. The update also refines agentic capabilities, especially for the Code Agent and Search Agent, leading to better tool integration and query handling. Benchmarks show small but notable gains, such as raising MMLU-Pro from 84.8 to 85.0, GPQA-Diamond from 80.1 to 80.7, and SWE Verified from 66.0 to 68.4, along with significant improvements in agent benchmarks like BrowseComp (30.0 → 38.5) and Terminal-bench (31.3 → 36.7). The model structure remains the same as DeepSeek-V3, ensuring compatibility with existing deployment methods, with updated inference demos provided for community use. -
23
DeiT (Data-efficient Image Transformers)
Official DeiT repository
DeiT (Data-efficient Image Transformers) shows that Vision Transformers can be trained competitively on ImageNet-1k without external data by using strong training recipes and knowledge distillation. Its key idea is a specialized distillation strategy—including a learnable “distillation token”—that lets a transformer learn effectively from a CNN or transformer teacher on modest-scale datasets. The project provides compact ViT variants (Tiny/Small/Base) that achieve excellent accuracy–throughput trade-offs, making transformers practical beyond massive pretraining regimes. Training involves carefully tuned augmentations, regularization, and optimization schedules to stabilize learning and improve sample efficiency. The repo offers pretrained checkpoints, reference scripts, and ablation studies that clarify which ingredients matter most for data-efficient ViT training. -
24
Denoiser
Real Time Speech Enhancement in the Waveform Domain (Interspeech 2020)
Denoiser is a real-time speech enhancement model operating directly on raw waveforms, designed to clean noisy audio while running efficiently on CPU. It uses a causal encoder-decoder architecture with skip connections, optimized with losses defined both in the time domain and frequency domain to better suppress noise while preserving speech. Unlike models that operate on spectrograms alone, this design enables lower latency and coherent waveform output. The implementation includes data augmentation techniques applied to the raw waveforms (e.g. noise mixing, reverberation) to improve model robustness and generalization to diverse noise types. The project supports both offline denoising (batch inference) and live audio processing (e.g. via loopback audio interfaces), making it practical for real-time use in calls or recording. The codebase includes training and evaluation scripts, configuration management via Hydra, and pretrained models on standard noise datasets. -
25
Depth Pro
Sharp Monocular Metric Depth in Less Than a Second
Depth Pro is a foundation model for zero-shot metric monocular depth estimation, producing sharp, high-frequency depth maps with absolute scale from a single image. Unlike many prior approaches, it does not require camera intrinsics or extra metadata, yet still outputs metric depth suitable for downstream 3D tasks. Apple highlights both accuracy and speed: the model can synthesize a ~2.25-megapixel depth map in around 0.3 seconds on a standard GPU, enabling near real-time applications. The repo and research page emphasize boundary fidelity and crisp geometry, addressing a common weakness in monocular depth where edges can blur. Community integrations (e.g., inference wrappers and UI nodes) have sprung up around the model, reflecting practical interest in video, AR, and generative pipelines. As a general-purpose monocular depth backbone, Depth Pro slots into 3D reconstruction, relighting, and scene understanding workflows that benefit from metric predictions.