Open Source AI Models - Page 6
Sort By:
-
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.
-
Gen AI apps are built with MongoDB AtlasMongoDB Atlas provides built-in vector search and a flexible document model so developers can build, scale, and run gen AI apps without stitching together multiple databases. From LLM integration to semantic search, Atlas simplifies your AI architecture—and it’s free to get started.
-
1
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. -
2
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. -
3
DiT (Diffusion Transformers)
Official PyTorch Implementation of "Scalable Diffusion Models"
DiT (Diffusion Transformer) is a powerful architecture that applies transformer-based modeling directly to diffusion generative processes for high-quality image synthesis. Unlike CNN-based diffusion models, DiT represents the diffusion process in the latent space and processes image tokens through transformer blocks with learned positional encodings, offering scalability and superior sample quality. The model architecture parallels large language models but for image tokens—each block refines noisy latent representations toward cleaner outputs through iterative denoising steps. DiT achieves strong results on benchmarks like ImageNet and LSUN while being architecturally simple and highly modular. It supports variable resolution, conditioning on class or text embeddings, and integration with latent autoencoders (like those used in Stable Diffusion). -
4
Dia-1.6B
Dia-1.6B generates lifelike English dialogue and vocal expressions
Dia-1.6B is a 1.6 billion parameter text-to-speech model by Nari Labs that generates high-fidelity dialogue directly from transcripts. Designed for realistic vocal performance, Dia supports expressive features like emotion, tone control, and non-verbal cues such as laughter, coughing, or sighs. The model accepts speaker conditioning through audio prompts, allowing limited voice cloning and speaker consistency across generations. It is optimized for English and built for real-time performance on enterprise GPUs, though CPU and quantized versions are planned. The format supports [S1]/[S2] tags to differentiate speakers and integrates easily into Python workflows. While not tuned to a specific voice, user-provided audio can guide output style. Licensed under Apache 2.0, Dia is intended for research and educational use, with explicit restrictions on misuse like identity mimicry or deceptive content. -
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
DreamCraft3D
Official implementation of DreamCraft3D
DreamCraft3D is DeepSeek’s generative 3D modeling framework / model family that likely extends their earlier 3D efforts (e.g. Shap-E or Point-E style models) with more capability, control, or expression. The name suggests a “dream crafting” metaphor—users probably supply textual or image prompts and generate 3D assets (point clouds, meshes, scenes). The repository includes model code, inference scripts, sample prompts, and possibly dataset preparation pipelines. It may integrate rendering or post-processing modules (e.g. mesh smoothing, texturing) to make the outputs more output-ready. Because 3D generation is hardware‐intensive, the repository likely also includes optimizations like quantization, pruning, or inference accelerations (e.g. using FlashMLA or DeepEP) to make the generation pipeline faster or more efficient. DreamCraft3D may also support style or attribute control (e.g. “make this object metallic,” “add textures”) via prompt conditioning or guides. -
6
FLUX.1-Krea-dev
Text-to-image model optimized for artistic quality and safe generation
FLUX.1-Krea-dev is a 12 billion parameter rectified flow transformer for text-to-image generation, developed by Black Forest Labs in collaboration with Krea. It delivers aesthetic, high-quality outputs focused on photography and visual coherence, making it a strong competitor to closed-source models. Trained using guidance distillation, it offers efficient inference while preserving creative fidelity. The model is distributed under a non-commercial license, with conditions to prevent misuse and support ethical AI development. FLUX.1-Krea-dev is available via Diffusers and ComfyUI, and integrates with the FluxPipeline for streamlined usage. Developers can use it for personal or scientific projects, but must comply with safety filters and content restrictions. Extensive pre- and post-training mitigations were applied to minimize risks like NSFW or abusive content generation. -
7
FastVLM
This repository contains the official implementation of FastVLM
FastVLM is an efficiency-focused vision-language modeling stack that introduces FastViTHD, a hybrid vision encoder engineered to emit fewer visual tokens and slash encoding time, especially for high-resolution images. Instead of elaborate pruning stages, the design trades off resolution and token count through input scaling, simplifying the pipeline while maintaining strong accuracy. Reported results highlight dramatic speedups in time-to-first-token and competitive quality versus contemporary open VLMs, including comparisons across small and larger variants. The repository documents model variants, showcases head-to-head numbers against known baselines, and explains how the encoder integrates with common LLM backbones. Apple’s research brief frames FastVLM as targeting real-time or latency-sensitive scenarios, where lowering visual token pressure is critical to interactive UX. In short, it’s a practical recipe to make VLMs fast without exotic token-selection heuristics. -
8
FastViT
This repository contains the official implementation of research
FastViT is an efficient vision backbone family that blends convolutional inductive biases with transformer capacity to deliver strong accuracy at mobile and real-time inference budgets. Its design pursues a favorable latency-accuracy Pareto curve, targeting edge devices and server scenarios where throughput and tail latency matter. The models use lightweight attention and carefully engineered blocks to minimize token mixing costs while preserving representation power. Training and inference recipes highlight straightforward integration into common vision tasks such as classification, detection, and segmentation. The codebase provides reference implementations and checkpoints that make it easy to evaluate or fine-tune on downstream datasets. In practice, FastViT offers drop-in backbones that reduce compute and memory pressure without exotic training tricks. -
9
FixRes
Reproduces results of "Fixing the train-test resolution discrepancy"
FixRes is a lightweight yet powerful training methodology for convolutional neural networks (CNNs) that addresses the common train-test resolution discrepancy problem in image classification. Developed by Facebook Research, FixRes improves model generalization by adjusting training and evaluation procedures to better align input resolutions used during different phases. The approach is simple but highly effective, requiring no architectural modifications and working across diverse CNN backbones such as ResNet, ResNeXt, PNASNet, and EfficientNet. FixRes demonstrates that a mismatch between training and testing resolutions often leads to suboptimal accuracy, and fine-tuning the classifier and batch normalization layers at higher test resolutions significantly enhances performance. The repository includes pretrained models, feature embeddings, and evaluation scripts corresponding to the experiments reported in the NeurIPS 2019 paper “Fixing the train-test resolution discrepancy.” -
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.
-
10
FlashMLA
FlashMLA: Efficient Multi-head Latent Attention Kernels
FlashMLA is a high-performance decoding kernel library designed especially for Multi-Head Latent Attention (MLA) workloads, targeting NVIDIA Hopper GPU architectures. It provides optimized kernels for MLA decoding, including support for variable-length sequences, helping reduce latency and increase throughput in model inference systems using that attention style. The library supports both BF16 and FP16 data types, and includes a paged KV cache implementation with a block size of 64 to efficiently manage memory during decoding. On very compute-bound settings, it can reach up to ~660 TFLOPS on H800 SXM5 hardware, while in memory-bound configurations it can push memory throughput to ~3000 GB/s. The team regularly updates it with performance improvements; for example, a 2025 update claims 5 % to 15 % gains on compute-bound workloads while maintaining API compatibility. -
11
Flow Matching
A PyTorch library for implementing flow matching algorithms
flow_matching is a PyTorch library implementing flow matching algorithms in both continuous and discrete settings, enabling generative modeling via matching vector fields rather than diffusion. The underlying idea is to parameterize a flow (a time-dependent vector field) that transports samples from a simple base distribution to a target distribution, and train via matching of flows without requiring score estimation or noisy corruption—this can lead to more efficient or stable generative training. The library supports both continuous-time flows (via differential equations) and discrete-time analogues, giving flexibility in design and tradeoffs. It provides examples across modalities (images, toy 2D distributions) to help users understand how to apply flow matching in practice. The codebase includes notebooks illustrating 2D flow matching, discrete flows, and Riemannian flow matching on curved manifolds (e.g. flat torus) for non-Euclidean support. -
12
GLIDE (Text2Im)
GLIDE: a diffusion-based text-conditional image synthesis model
glide-text2im is an open source implementation of OpenAI’s GLIDE model, which generates photorealistic images from natural language text prompts. It demonstrates how diffusion-based generative models can be conditioned on text to produce highly detailed and coherent visual outputs. The repository provides both model code and pretrained checkpoints, making it possible for researchers and developers to experiment with text-to-image synthesis. GLIDE includes advanced techniques such as classifier-free guidance, which improves the quality and alignment of generated images with the input text. The project also offers sampling scripts and utilities for exploring how diffusion models can be applied to multimodal tasks. As one of the early diffusion-based text-to-image systems, glide-text2im laid important groundwork for later advances in generative AI research. -
13
GLM-4
GLM-4 series: Open Multilingual Multimodal Chat LMs
GLM-4 is a family of open models from ZhipuAI that spans base, chat, and reasoning variants at both 32B and 9B scales, with long-context support and practical local-deployment options. The GLM-4-32B-0414 models are trained on ~15T high-quality data (including substantial synthetic reasoning data), then post-trained with preference alignment, rejection sampling, and reinforcement learning to improve instruction following, coding, function calling, and agent-style behaviors. The GLM-Z1-32B-0414 line adds deeper mathematical, coding, and logical reasoning via extended reinforcement learning and pairwise ranking feedback, while GLM-Z1-Rumination-32B-0414 introduces a “rumination” mode that performs longer, tool-using deep research for complex, open-ended tasks. A lightweight GLM-Z1-9B-0414 brings many of these techniques to a smaller model, targeting strong reasoning under tight resource budgets. -
14
GLM-4-32B-0414
Open Multilingual Multimodal Chat LMs
GLM-4-32B-0414 is a powerful open-source large language model featuring 32 billion parameters, designed to deliver performance comparable to leading models like OpenAI’s GPT series. It supports multilingual and multimodal chat capabilities with an extensive 32K token context length, making it ideal for dialogue, reasoning, and complex task completion. The model is pre-trained on 15 trillion tokens of high-quality data, including substantial synthetic reasoning datasets, and further enhanced with reinforcement learning and human preference alignment for improved instruction-following and function calling. Variants like GLM-Z1-32B-0414 offer deep reasoning and advanced mathematical problem-solving, while GLM-Z1-Rumination-32B-0414 specializes in long-form, complex research-style writing using scaled reinforcement learning and external search tools. Despite its large capacity, the model supports user-friendly local deployment and efficient fine-tuning with available scripts. -
15
GLM-4-Voice
GLM-4-Voice | End-to-End Chinese-English Conversational Model
GLM-4-Voice is an open-source speech-enabled model from ZhipuAI, extending the GLM-4 family into the audio domain. It integrates advanced voice recognition and generation with the multimodal reasoning capabilities of GLM-4, enabling smooth natural interaction via spoken input and output. The model supports real-time speech-to-text transcription, spoken dialogue understanding, and text-to-speech synthesis, making it suitable for conversational AI, virtual assistants, and accessibility applications. GLM-4-Voice builds upon the bilingual strengths of the GLM architecture, supporting both Chinese and English, and is designed to handle long-form conversations with context retention. The repository provides model weights, inference demos, and setup instructions for deploying speech-enabled AI systems. -
16
GLM-4.5-Air
Compact hybrid reasoning language model for intelligent responses
GLM-4.5-Air is a multilingual large language model with 106 billion total parameters and 12 billion active parameters, designed for conversational AI and intelligent agents. It is part of the GLM-4.5 family developed by Zhipu AI, offering hybrid reasoning capabilities via two modes: a thinking mode for complex reasoning and tool use, and a non-thinking mode for immediate responses. The model is optimized for efficiency and deployment, delivering strong results across 12 industry benchmarks, with a composite score of 59.8. GLM-4.5-Air supports both English and Chinese, and is suitable for tasks involving text generation, coding, reasoning, and tool calling. Open-sourced under the MIT license, it is commercially usable and integrates with transformers, vLLM, and SGLang inference frameworks. It includes FP8 variants for faster inference and reduced memory requirements. Despite its smaller size compared to full GLM-4.5, GLM-4.5-Air maintains high performance. -
17
GLM-V
GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning
GLM-V is an open-source vision-language model (VLM) series from ZhipuAI that extends the GLM foundation models into multimodal reasoning and perception. The repository provides both GLM-4.5V and GLM-4.1V models, designed to advance beyond basic perception toward higher-level reasoning, long-context understanding, and agent-based applications. GLM-4.5V builds on the flagship GLM-4.5-Air foundation (106B parameters, 12B active), achieving state-of-the-art results on 42 benchmarks across image, video, document, GUI, and grounding tasks. It introduces hybrid training for broad-spectrum reasoning and a Thinking Mode switch to balance speed and depth of reasoning. GLM-4.1V-9B-Thinking incorporates reinforcement learning with curriculum sampling (RLCS) and Chain-of-Thought reasoning, outperforming models much larger in scale (e.g., Qwen-2.5-VL-72B) across many benchmarks. -
18
GPT-NeoX
Implementation of model parallel autoregressive transformers on GPUs
This repository records EleutherAI's library for training large-scale language models on GPUs. Our current framework is based on NVIDIA's Megatron Language Model and has been augmented with techniques from DeepSpeed as well as some novel optimizations. We aim to make this repo a centralized and accessible place to gather techniques for training large-scale autoregressive language models, and accelerate research into large-scale training. For those looking for a TPU-centric codebase, we recommend Mesh Transformer JAX. If you are not looking to train models with billions of parameters from scratch, this is likely the wrong library to use. For generic inference needs, we recommend you use the Hugging Face transformers library instead which supports GPT-NeoX models. -
19
Gemma in PyTorch
The official PyTorch implementation of Google's Gemma models
gemma_pytorch provides the official PyTorch reference for running and fine-tuning Google’s Gemma family of open models. It includes model definitions, configuration files, and loading utilities for multiple parameter scales, enabling quick evaluation and downstream adaptation. The repository demonstrates text generation pipelines, tokenizer setup, quantization paths, and adapters for low-rank or parameter-efficient fine-tuning. Example notebooks walk through instruction tuning and evaluation so teams can benchmark and iterate rapidly. The code is organized to be legible and hackable, exposing attention blocks, positional encodings, and head configurations. With standard PyTorch abstractions, it integrates easily into existing training loops, loggers, and evaluation harnesses. -
20
Granite 3.0 Language Models
New set of lightweight state-of-the-art, open foundation models
This repository introduces Granite 3.0 language models as lightweight, state-of-the-art open foundation models built to natively support multilinguality, coding, reasoning, and tool usage. A central goal is efficient deployment, including the potential to run on constrained compute resources while remaining useful for a broad span of enterprise tasks. The repo positions the models for both research and commercial use under an Apache-2.0 license, signaling permissive adoption paths. Documentation highlights the capability mix (reasoning, tool use, code) and points to model artifacts and guidance for evaluation. Activity on the project shows an evolving codebase with open pull requests and standard GitHub project structure for issues and security visibility. In practice, this is a hub for acquiring Granite 3.0 variants and understanding how to integrate them into applications. -
21
Granite Code Models
A Family of Open Foundation Models for Code Intelligence
Granite Code Models are IBM’s open-source, decoder-only models tailored for code tasks such as fixing bugs, explaining and documenting code, and modernizing codebases. Trained on code from 116 programming languages, the family targets strong performance across diverse benchmarks while remaining accessible to the community. The repository introduces the model lineup, intended uses, and evaluation highlights, and it complements IBM’s broader Granite initiative spanning multiple modalities. IBM’s research blog details the motivation for opening these models and points developers to downloads, papers, and hosting options. Together, the materials position Granite Code as enterprise-friendly, permissively licensed models for practical software engineering assistance. They slot into the larger Granite ecosystem that includes language and time-series models, community cookbooks, and production guidance. -
22
Grok-2.5
Large-scale xAI model for local inference with SGLang, Grok-2.5
Grok-2.5 is a large-scale AI model developed and released by xAI in 2024, made available through Hugging Face for research and experimentation. The model is distributed as raw weights that require specialized infrastructure to run, rather than being hosted by inference providers. To use it, users must download over 500 GB of files and set them up locally with the SGLang inference engine. Grok-2.5 supports advanced inference with multi-GPU configurations, requiring at least 8 GPUs with more than 40 GB of memory each for optimal performance. It integrates with the SGLang framework to enable serving, testing, and chat-style interactions. The model comes with a post-training architecture and requires the correct chat template to function properly. It is released under the Grok 2 Community License Agreement, encouraging community experimentation and responsible use. -
23
Hermes 4
Hermes 4 FP8: hybrid reasoning Llama-3.1-405B model by Nous Research
Hermes 4 405B FP8 is a cutting-edge large language model developed by Nous Research, built on Llama-3.1-405B and optimized for frontier reasoning and alignment. It introduces a hybrid reasoning mode with explicit <think> segments, enabling the model to deliberate deeply when needed and switch to faster responses when desired. Post-training improvements include a vastly expanded corpus with ~60B tokens, boosting performance across math, code, STEM, logic, creativity, and structured outputs. The model is designed for schema adherence, producing valid JSON and repairing malformed outputs, making it highly suitable for tool use and function calling. Hermes 4 is engineered for superior steerability with reduced refusal rates, aligning responses to user values while preserving assistant quality. It achieves state-of-the-art results on RefusalBench, outperforming both closed and open models in balancing helpfulness with adaptability. -
24
Hunyuan-A13B-Instruct
Efficient 13B MoE language model with long context and reasoning modes
Hunyuan-A13B-Instruct is a powerful instruction-tuned large language model developed by Tencent using a fine-grained Mixture-of-Experts (MoE) architecture. While the total model includes 80 billion parameters, only 13 billion are active per forward pass, making it highly efficient while maintaining strong performance across benchmarks. It supports up to 256K context tokens, advanced reasoning (CoT) abilities, and agent-based workflows with tool parsing. The model offers both fast and slow thinking modes, letting users trade off speed for deeper reasoning. It excels in mathematics, science, coding, and multi-turn conversation tasks, rivaling or outperforming larger models in several areas. Deployment is supported via TensorRT-LLM, vLLM, and SGLang, with Docker images and integration guides provided. Open-source under a custom license, it's ideal for researchers and developers seeking scalable, high-context AI capabilities with optimized inference. -
25
Hunyuan-MT-7B
Tencent’s 36-language state-of-the-art translation model
Hunyuan-MT-7B is a large-scale multilingual translation model developed by Tencent, designed to deliver state-of-the-art translation quality across 36 languages, including several Chinese ethnic minority languages. It forms part of the Hunyuan Translation Model family, alongside Hunyuan-MT-Chimera, which ensembles outputs for even higher accuracy. Trained with a comprehensive framework spanning pretraining, cross-lingual pretraining, supervised fine-tuning, enhancement, and ensemble refinement, the model achieves competitive results against systems of similar or larger scale. At WMT25, Hunyuan-MT secured first place in 30 of the 31 translation categories it entered, highlighting its reliability in real-world multilingual use. The model offers flexible deployment options with variants such as FP8 quantization and ensemble configurations. It supports natural prompts for both Chinese translations, making it easy to integrate into translation workflows.