Open Source AI Models - Page 10
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.
-
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.
-
1
bge-small-en-v1.5
Compact English sentence embedding model for semantic search tasks
BAAI/bge-small-en-v1.5 is a lightweight English sentence embedding model developed by the Beijing Academy of Artificial Intelligence (BAAI) as part of the BGE (BAAI General Embedding) series. Designed for dense retrieval, semantic search, and similarity tasks, it produces 384-dimensional embeddings that can be used to compare and rank sentences or passages. This version (v1.5) improves similarity distribution, enhancing performance without the need for special query instructions. The model is optimized for speed and efficiency, making it suitable for resource-constrained environments. It is compatible with popular libraries such as FlagEmbedding, Sentence-Transformers, and Hugging Face Transformers. The model achieves competitive results on the MTEB benchmark, especially in retrieval and classification tasks. With only 33.4M parameters, it provides a strong balance of accuracy and performance for English-only use cases. -
2
fairseq-lua
Facebook AI Research Sequence-to-Sequence Toolkit
fairseq-lua is the original Lua/Torch7 version of Facebook AI Research’s sequence modeling toolkit, designed for neural machine translation (NMT) and sequence generation. It introduced early attention-based architectures and training pipelines that later evolved into the modern PyTorch-based fairseq. The framework implements sequence-to-sequence models with attention, beam search decoding, and distributed training, providing a research platform for exploring translation, summarization, and language modeling. Its modular design made it easy to prototype new architectures by modifying encoders, decoders, or attention mechanisms. Although now deprecated in favor of the PyTorch rewrite, fairseq-lua played a key role in advancing large-scale NMT systems, such as early versions of Facebook’s production translation models. It remains an important historical reference for neural sequence learning frameworks. -
3
fairseq2
FAIR Sequence Modeling Toolkit 2
fairseq2 is a modern, modular sequence modeling framework developed by Meta AI Research as a complete redesign of the original fairseq library. Built from the ground up for scalability, composability, and research flexibility, fairseq2 supports a broad range of language, speech, and multimodal content generation tasks, including instruction fine-tuning, reinforcement learning from human feedback (RLHF), and large-scale multilingual modeling. Unlike the original fairseq—which evolved into a large, monolithic codebase—fairseq2 introduces a clean, plugin-oriented architecture designed for long-term maintainability and rapid experimentation. It supports multi-GPU and multi-node distributed training using DDP, FSDP, and tensor parallelism, capable of scaling up to 70B+ parameter models. The framework integrates seamlessly with PyTorch 2.x features such as torch.compile, Fully Sharded Data Parallel (FSDP), and modern configuration management. -
4
fashion-clip
CLIP model fine-tuned for zero-shot fashion product classification
FashionCLIP is a domain-adapted CLIP model fine-tuned specifically for the fashion industry, enabling zero-shot classification and retrieval of fashion products. Developed by Patrick John Chia and collaborators, it builds on the CLIP ViT-B/32 architecture and was trained on over 800K image-text pairs from the Farfetch dataset. The model learns to align product images and descriptive text using contrastive learning, enabling it to perform well across various fashion-related tasks without additional supervision. FashionCLIP 2.0, the latest version, uses the laion/CLIP-ViT-B-32-laion2B-s34B-b79K checkpoint for improved accuracy, achieving better F1 scores across multiple benchmarks compared to earlier versions. It supports multilingual fashion queries and works best with clean, product-style images against white backgrounds. The model can be used for product search, recommendation systems, or visual tagging in e-commerce platforms. -
MongoDB Atlas runs apps anywhereMongoDB Atlas gives you the freedom to build and run modern applications anywhere—across AWS, Azure, and Google Cloud. With global availability in over 115 regions, Atlas lets you deploy close to your users, meet compliance needs, and scale with confidence across any geography.
-
5
gpt-oss-120b
OpenAI’s open-weight 120B model optimized for reasoning and tooling
GPT-OSS-120B is a powerful open-weight language model by OpenAI, optimized for high-level reasoning, tool use, and agentic tasks. With 117B total parameters and 5.1B active parameters, it’s designed to fit on a single H100 GPU using native MXFP4 quantization. The model supports fine-tuning, chain-of-thought reasoning, and structured outputs, making it ideal for complex workflows. It operates in OpenAI’s Harmony response format and can be deployed via Transformers, vLLM, Ollama, LM Studio, and PyTorch. Developers can control the reasoning level (low, medium, high) to balance speed and depth depending on the task. Released under the Apache 2.0 license, it enables both commercial and research applications. The model supports function calling, web browsing, and code execution, streamlining intelligent agent development. -
6
gpt-oss-20b
OpenAI’s compact 20B open model for fast, agentic, and local use
GPT-OSS-20B is OpenAI’s smaller, open-weight language model optimized for low-latency, agentic tasks, and local deployment. With 21B total parameters and 3.6B active parameters (MoE), it fits within 16GB of memory thanks to native MXFP4 quantization. Designed for high-performance reasoning, it supports Harmony response format, function calling, web browsing, and code execution. Like its larger sibling (gpt-oss-120b), it offers adjustable reasoning depth and full chain-of-thought visibility for better interpretability. It’s released under a permissive Apache 2.0 license, allowing unrestricted commercial and research use. GPT-OSS-20B is compatible with Transformers, vLLM, Ollama, PyTorch, and other tools. It is ideal for developers building lightweight AI agents or experimenting with fine-tuning on consumer-grade hardware. -
7
granite-timeseries-ttm-r2
Tiny pre-trained IBM model for multivariate time series forecasting
granite-timeseries-ttm-r2 is part of IBM’s TinyTimeMixers (TTM) series—compact, pre-trained models for multivariate time series forecasting. Unlike massive foundation models, TTM models are designed to be lightweight yet powerful, with only ~805K parameters, enabling high performance even on CPU or single-GPU machines. The r2 version is pre-trained on ~700M samples (r2.1 expands to ~1B), delivering up to 15% better accuracy than the r1 version. TTM supports both zero-shot and fine-tuned forecasting, handling minutely, hourly, daily, and weekly resolutions. It can integrate exogenous variables, static categorical features, and perform channel-mixing for richer multivariate forecasting. The get_model() utility makes it easy to auto-select the best TTM model for specific context and prediction lengths. These models significantly outperform benchmarks like Chronos, GPT4TS, and Moirai while demanding a fraction of the compute. -
8
layoutlm-base-uncased
Multimodal Transformer for document image understanding and layout
layoutlm-base-uncased is a multimodal transformer model developed by Microsoft for document image understanding tasks. It incorporates both text and layout (position) features to effectively process structured documents like forms, invoices, and receipts. This base version has 113 million parameters and is pre-trained on 11 million documents from the IIT-CDIP dataset. LayoutLM enables better performance in tasks where the spatial arrangement of text plays a crucial role. The model uses a standard BERT-like architecture but enriches input with 2D positional embeddings. It achieves state-of-the-art results in form understanding and information extraction benchmarks. This model is particularly useful for document AI applications like document classification, question answering, and named entity recognition. -
9
minGPT
A minimal PyTorch re-implementation of the OpenAI GPT
minGPT is a minimalist, educational re-implementation of the GPT (Generative Pretrained Transformer) architecture built in PyTorch, designed by Andrej Karpathy to expose the core structure of a transformer-based language model in as few lines of code as possible. It strips away extraneous bells and whistles, aiming to show how a sequence of token indices is fed into a stack of transformer blocks and then decoded into the next token probabilities, with both training and inference supported. Because the whole model is around 300 lines of code, users can follow each step—from embedding lookup, positional encodings, multi-head attention, feed-forward layers, to output heads—and thus demystify how GPT-style models work beneath the surface. It provides a practical sandbox for experimentation, letting learners tweak the architecture, dataset, or training loop without being overwhelmed by framework abstraction. -
Boon: The Agile Referral Hiring PlatformBoon's agile referral platform expands your recruiting power through AI, automation, integrations, and gamification.
-
10
mms-300m-1130-forced-aligner
CTC-based forced aligner for audio-text in 158 languages
mms-300m-1130-forced-aligner is a multilingual forced alignment model based on Meta’s MMS-300M wav2vec2 checkpoint, adapted for Hugging Face’s Transformers library. It supports forced alignment between audio and corresponding text across 158 languages, offering broad multilingual coverage. The model enables accurate word- or phoneme-level timestamping using Connectionist Temporal Classification (CTC) emissions. Unlike other tools, it provides significant memory efficiency compared to the TorchAudio forced alignment API. Users can integrate it easily through the Python package ctc-forced-aligner, and it supports GPU acceleration via PyTorch. The alignment pipeline includes audio processing, emission generation, tokenization, and span detection, making it suitable for speech analysis, transcription syncing, and dataset creation. This model is especially useful for researchers and developers working with low-resource languages or building multilingual speech systems. -
11
pyllama
LLaMA: Open and Efficient Foundation Language Models
📢 pyllama is a hacked version of LLaMA based on original Facebook's implementation but more convenient to run in a Single consumer grade GPU. -
12
roberta-base
Robust BERT-based model for English with improved MLM training
roberta-base is a robustly optimized variant of BERT, pretrained on a significantly larger corpus of English text using dynamic masked language modeling. Developed by Facebook AI, RoBERTa improves on BERT by removing the Next Sentence Prediction objective, using longer training, larger batches, and more data, including BookCorpus, English Wikipedia, CC-News, OpenWebText, and Stories. It captures contextual representations of language by masking 15% of input tokens and predicting them. RoBERTa is designed to be fine-tuned for a wide range of NLP tasks such as classification, QA, and sequence labeling, achieving strong performance on the GLUE benchmark and other downstream applications. -
13
rwkv.cpp
INT4/INT5/INT8 and FP16 inference on CPU for RWKV language model
Besides the usual FP32, it supports FP16, quantized INT4, INT5 and INT8 inference. This project is focused on CPU, but cuBLAS is also supported. RWKV is a novel large language model architecture, with the largest model in the family having 14B parameters. In contrast to Transformer with O(n^2) attention, RWKV requires only state from the previous step to calculate logits. This makes RWKV very CPU-friendly on large context lengths. -
14
t5-base
Flexible text-to-text transformer model for multilingual NLP tasks
t5-base is a pre-trained transformer model from Google’s T5 (Text-To-Text Transfer Transformer) family that reframes all NLP tasks into a unified text-to-text format. With 220 million parameters, it can handle a wide range of tasks, including translation, summarization, question answering, and classification. Unlike traditional models like BERT, which output class labels or spans, T5 always generates text outputs. It was trained on the C4 dataset, along with a variety of supervised NLP benchmarks, using both unsupervised denoising and supervised objectives. The model supports multiple languages, including English, French, Romanian, and German. Its flexible architecture and consistent input/output format simplify model reuse and transfer learning across different NLP tasks. T5-base achieves competitive performance across 24 language understanding tasks, as documented in its research paper. -
15
t5-small
T5-Small: Lightweight text-to-text transformer for NLP tasks
T5-Small is a lightweight variant of the Text-To-Text Transfer Transformer (T5), designed to handle a wide range of NLP tasks using a unified text-to-text approach. Developed by researchers at Google, this model reframes all tasks—such as translation, summarization, classification, and question answering—into the format of input and output as plain text strings. With only 60 million parameters, T5-Small is compact and suitable for fast inference or deployment in constrained environments. It was pretrained on the C4 dataset using both unsupervised denoising and supervised learning on tasks like sentiment analysis, NLI, and QA. Despite its size, it performs competitively across 24 NLP benchmarks, making it a strong candidate for prototyping and fine-tuning. T5-Small is compatible with major deep learning frameworks including PyTorch, TensorFlow, JAX, and ONNX. The model is open-source under the Apache 2.0 license and has wide support across Hugging Face's ecosystem. -
16
unidepth-v2-vitl14
Metric monocular depth estimation (vision model)
Estimates absolute (metric) depth from single RGB images, along with camera intrinsics and uncertainty. Designed to generalize across domains (zero-shot) using a self‑prompting camera module and pseudo-spherical prediction space. -
17
wav2vec2-large-xlsr-53-portuguese
Portuguese ASR model fine-tuned on XLSR-53 for 16kHz audio input
wav2vec2-large-xlsr-53-portuguese is an automatic speech recognition (ASR) model fine-tuned on Portuguese using the Common Voice 6.1 dataset. It is based on Facebook’s wav2vec2-large-xlsr-53, a multilingual self-supervised learning model, and is optimized to transcribe Portuguese speech sampled at 16kHz. The model performs well without a language model, though adding one can improve word error rate (WER) and character error rate (CER). It achieves a WER of 11.3% (or 9.01% with LM) on Common Voice test data, demonstrating high accuracy for a single-language ASR model. Inference can be done using HuggingSound or via a custom PyTorch script using Hugging Face Transformers and Librosa. Training scripts and evaluation methods are open source and available on GitHub. It is released under the Apache 2.0 license and intended for ASR tasks in Brazilian Portuguese. -
18
wav2vec2-large-xlsr-53-russian
Russian ASR model fine-tuned on Common Voice and CSS10 datasets
wav2vec2-large-xlsr-53-russian is a fine-tuned automatic speech recognition (ASR) model based on Facebook’s wav2vec2-large-xlsr-53 and optimized for Russian. It was trained using Mozilla’s Common Voice 6.1 and CSS10 datasets to recognize Russian speech with high accuracy. The model operates best with audio sampled at 16kHz and can transcribe Russian speech directly without a language model. It achieves a Word Error Rate (WER) of 13.3% and Character Error Rate (CER) of 2.88% on the Common Voice test set, with even better results when used with a language model. The model supports both PyTorch and JAX and is compatible with the Hugging Face Transformers and HuggingSound libraries. It is ideal for Russian voice transcription tasks in research, accessibility, and interface development. The training was made possible with compute support from OVHcloud, and the training scripts are publicly available for replication. -
19
xFormers
Hackable and optimized Transformers building blocks
xformers is a modular, performance-oriented library of transformer building blocks, designed to allow researchers and engineers to compose, experiment, and optimize transformer architectures more flexibly than monolithic frameworks. It abstracts components like attention layers, feedforward modules, normalization, and positional encoding, so you can mix and match or swap optimized kernels easily. One of its key goals is efficient attention: it supports dense, sparse, low-rank, and approximate attention mechanisms (e.g. FlashAttention, Linformer, Performer) via interchangeable modules. The library includes memory-efficient operator implementations in both Python and optimized C++/CUDA, ensuring that performance isn’t sacrificed for modularity. It also integrates with PyTorch seamlessly so you can drop in its blocks to existing models, replace default attention layers, or build new architectures from scratch. xformers includes training, deployment, and memory profiling tools.