Open Source AI Models
Sort By:
AI Models
Browse free open source AI Models and projects for Windows and BSD below. Use the toggles on the left to filter open source AI Models by OS, license, language, programming language, and project status.
-
Full-stack observability with actually useful AI | Grafana CloudBuilt on open standards like Prometheus and OpenTelemetry, Grafana Cloud includes Kubernetes Monitoring, Application Observability, Incident Response, plus the AI-powered Grafana Assistant. Get started with our generous free tier today.
-
Fully Managed MySQL, PostgreSQL, and SQL ServerCloud SQL handles your database ops end to end, so you can focus on your app.
-
1
DeepSeek R1
Open-source, high-performance AI model with advanced reasoning
DeepSeek-R1 is an open-source large language model developed by DeepSeek, designed to excel in complex reasoning tasks across domains such as mathematics, coding, and language. DeepSeek R1 offers unrestricted access for both commercial and academic use. The model employs a Mixture of Experts (MoE) architecture, comprising 671 billion total parameters with 37 billion active parameters per token, and supports a context length of up to 128,000 tokens. DeepSeek-R1's training regimen uniquely integrates large-scale reinforcement learning (RL) without relying on supervised fine-tuning, enabling the model to develop advanced reasoning capabilities. This approach has resulted in performance comparable to leading models like OpenAI's o1, while maintaining cost-efficiency. To further support the research community, DeepSeek has released distilled versions of the model based on architectures such as LLaMA and Qwen. -
2
GLM-5
From Vibe Coding to Agentic Engineering
GLM-5 is a next-generation open-source large language model (LLM) developed by the Z .ai team under the zai-org organization that pushes the boundaries of reasoning, coding, and long-horizon agentic intelligence. Building on earlier GLM series models, GLM-5 dramatically scales the parameter count (to roughly 744 billion) and expands pre-training data to significantly improve performance on complex tasks such as multi-step reasoning, software engineering workflows, and agent orchestration compared to its predecessors like GLM-4.5. It incorporates innovations like DeepSeek Sparse Attention (DSA) to preserve massive context windows while reducing deployment costs and supporting long context processing, which is crucial for detailed plans and agent tasks. -
3
DeepSeek-V3
Powerful AI language model (MoE) optimized for efficiency/performance
DeepSeek-V3 is a robust Mixture-of-Experts (MoE) language model developed by DeepSeek, featuring a total of 671 billion parameters, with 37 billion activated per token. It employs Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture to enhance computational efficiency. The model introduces an auxiliary-loss-free load balancing strategy and a multi-token prediction training objective to boost performance. Trained on 14.8 trillion diverse, high-quality tokens, DeepSeek-V3 underwent supervised fine-tuning and reinforcement learning to fully realize its capabilities. Evaluations indicate that it outperforms other open-source models and rivals leading closed-source models, achieving this with a training duration of 55 days on 2,048 Nvidia H800 GPUs, costing approximately $5.58 million. -
4
LTX-2.3
Official Python inference and LoRA trainer package
LTX-2.3 is an open-source multimodal artificial intelligence foundation model developed by Lightricks for generating synchronized video and audio from prompts or other inputs. Unlike most earlier video generation systems that only produced silent clips, LTX-2 combines video and audio generation in a unified architecture capable of producing coherent audiovisual scenes. The model uses a diffusion-transformer-based architecture designed to generate high-fidelity visual frames while simultaneously producing corresponding audio elements such as speech, music, ambient sound, or effects. This unified approach allows creators to generate complete multimedia sequences where motion, timing, and sound are aligned automatically. LTX-2 is designed for both research and production workflows and can generate high-resolution video clips with precise control over structure, motion, and camera behavior. -
AI-generated apps that pass security reviewRetool lets you generate dashboards, admin panels, and workflows directly on your data. Type something like “Build me a revenue dashboard on my Stripe data” and get a working app with security, permissions, and compliance built in from day one. Whether on our cloud or self-hosted, create the internal software your team needs without compromising enterprise standards or control.
-
5
SAM 3
Code for running inference and finetuning with SAM 3 model
SAM 3 (Segment Anything Model 3) is a unified foundation model for promptable segmentation in both images and videos, capable of detecting, segmenting, and tracking objects. It accepts both text prompts (open-vocabulary concepts like “red car” or “goalkeeper in white”) and visual prompts (points, boxes, masks) and returns high-quality masks, boxes, and scores for the requested concepts. Compared with SAM 2, SAM 3 introduces the ability to exhaustively segment all instances of an open-vocabulary concept specified by a short phrase or exemplars, scaling to a vastly larger set of categories than traditional closed-set models. This capability is grounded in a new data engine that automatically annotated over four million unique concepts, producing a massive open-vocabulary segmentation dataset and enabling the model to achieve 75–80% of human performance on the SA-CO benchmark, which itself spans 270K unique concepts. -
6
llama.cpp
Port of Facebook's LLaMA model in C/C++
The llama.cpp project enables the inference of Meta's LLaMA model (and other models) in pure C/C++ without requiring a Python runtime. It is designed for efficient and fast model execution, offering easy integration for applications needing LLM-based capabilities. The repository focuses on providing a highly optimized and portable implementation for running large language models directly within C/C++ environments. -
7
GLM-4.7
Advanced language and coding AI model
GLM-4.7 is an advanced agent-oriented large language model designed as a high-performance coding and reasoning partner. It delivers significant gains over GLM-4.6 in multilingual agentic coding, terminal-based workflows, and real-world developer benchmarks such as SWE-bench and Terminal Bench 2.0. The model introduces stronger “thinking before acting” behavior, improving stability and accuracy in complex agent frameworks like Claude Code, Cline, and Roo Code. GLM-4.7 also advances “vibe coding,” producing cleaner, more modern UIs, better-structured webpages, and visually improved slide layouts. Its tool-use capabilities are substantially enhanced, with notable improvements in browsing, search, and tool-integrated reasoning tasks. Overall, GLM-4.7 shows broad performance upgrades across coding, reasoning, chat, creative writing, and role-play scenarios. -
8
GLM-4.6
Agentic, Reasoning, and Coding (ARC) foundation models
GLM-4.6 is the latest iteration of Zhipu AI’s foundation model, delivering significant advancements over GLM-4.5. It introduces an extended 200K token context window, enabling more sophisticated long-context reasoning and agentic workflows. The model achieves superior coding performance, excelling in benchmarks and practical coding assistants such as Claude Code, Cline, Roo Code, and Kilo Code. Its reasoning capabilities have been strengthened, including improved tool usage during inference and more effective integration within agent frameworks. GLM-4.6 also enhances writing quality, producing outputs that better align with human preferences and role-playing scenarios. Benchmark evaluations demonstrate that it not only outperforms GLM-4.5 but also rivals leading global models such as DeepSeek-V3.1-Terminus and Claude Sonnet 4. -
9
LTX-2
Python inference and LoRA trainer package for the LTX-2 audio–video
LTX-2 is a powerful, open-source toolkit developed by Lightricks that provides a modular, high-performance base for building real-time graphics and visual effects applications. It is architected to give developers low-level control over rendering pipelines, GPU resource management, shader orchestration, and cross-platform abstractions so they can craft visually compelling experiences without starting from scratch. Beyond basic rendering scaffolding, LTX-2 includes optimized math libraries, resource loaders, utilities for texture and buffer handling, and integration points for native event loops and input systems. The framework targets both interactive graphical applications and media-rich experiences, making it a solid foundation for games, creative tools, or visualization systems that demand both performance and flexibility. While being low-level, it also provides sensible defaults and helper abstractions that reduce boilerplate and help teams maintain clear, maintainable code. -
Gemini 3 and 200+ AI Models on One PlatformBuild generative AI apps with Vertex AI. Switch between models without switching platforms.
-
10
FLUX.2
Official inference repo for FLUX.2 models
FLUX.2 is a state-of-the-art open-weight image generation and editing model released by Black Forest Labs aimed at bridging the gap between research-grade capabilities and production-ready workflows. The model offers both text-to-image generation and powerful image editing, including editing of multiple reference images, with fidelity, consistency, and realism that push the limits of what open-source generative models have achieved. It supports high-resolution output (up to ~4 megapixels), which allows for photography-quality images, detailed product shots, infographics or UI mockups rather than just low-resolution drafts. FLUX.2 is built with a modern architecture (a flow-matching transformer + a revamped VAE + a strong vision-language encoder), enabling strong prompt adherence, correct rendering of text/typography in images, reliable lighting, layout, and physical realism, and consistent style/character/product identity across multiple generations or edits. -
11
GLM-4.5
GLM-4.5: Open-source LLM for intelligent agents by Z.ai
GLM-4.5 is a cutting-edge open-source large language model designed by Z.ai for intelligent agent applications. The flagship GLM-4.5 model has 355 billion total parameters with 32 billion active parameters, while the compact GLM-4.5-Air version offers 106 billion total parameters and 12 billion active parameters. Both models unify reasoning, coding, and intelligent agent capabilities, providing two modes: a thinking mode for complex reasoning and tool usage, and a non-thinking mode for immediate responses. They are released under the MIT license, allowing commercial use and secondary development. GLM-4.5 achieves strong performance on 12 industry-standard benchmarks, ranking 3rd overall, while GLM-4.5-Air balances competitive results with greater efficiency. The models support FP8 and BF16 precision, and can handle very large context windows of up to 128K tokens. Flexible inference is supported through frameworks like vLLM and SGLang with tool-call and reasoning parsers included. -
12
Z-Image
Image generation model with single-stream diffusion transformer
Z-Image is an efficient, open-source image generation foundation model built to make high-quality image synthesis more accessible. With just 6 billion parameters — far fewer than many large-scale models — it uses a novel “single-stream diffusion Transformer” architecture to deliver photorealistic image generation, demonstrating that excellence does not always require extremely large model sizes. The project includes several variants: Z-Image-Turbo, a distilled version optimized for speed and low resource consumption; Z-Image-Base, the full-capacity foundation model; and Z-Image-Edit, fine-tuned for image editing tasks. Despite its compact size, Z-Image produces outputs that closely rival those from much larger models — including strong rendering of bilingual (English and Chinese) text inside images, accurate prompt adherence, and good layout and composition. -
13
Qwen3.5
Qwen3.5 is the large language model series developed by Qwen team
Qwen3.5 is part of Alibaba’s Qwen family of large language and multimodal foundation models, designed to power advanced AI applications such as chatbots, coding assistants, and autonomous agents. The project represents a significant step toward “agentic AI,” meaning models that can reason through multi-step tasks and interact with external tools or environments rather than only generating text. Qwen3.5 builds on earlier Qwen generations by improving multilingual understanding, reasoning ability, and efficiency, while also introducing native multimodal capabilities that allow the model to work with both language and visual inputs. Architecturally, the system leverages modern large-scale training techniques and mixture-of-experts style efficiency so that very large parameter counts can be used while keeping inference practical. -
14
DeepSeek Coder V2
DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models
DeepSeek-Coder-V2 is the version-2 iteration of DeepSeek’s code generation models, refining the original DeepSeek-Coder line with improved architecture, training strategies, and benchmark performance. While the V1 models already targeted strong code understanding and generation, V2 appears to push further in both multilingual support and reasoning in code, likely via architectural enhancements or additional training objectives. The repository provides updated model weights, evaluation results on benchmarks (e.g. HumanEval, MultiPL-E, APPS), and new inference/serving scripts. Compared to the original, DeepSeek-Coder-V2 likely incorporates improved context management, caching strategies, or enhanced infilling capabilities. The project aims to provide a more performant and reliable open-source alternative to closed-source code models, optimized for practical usage in code completion, infilling, and code understanding across English and Chinese codebases. -
15
Stable Diffusion
High-Resolution Image Synthesis with Latent Diffusion Models
Stable Diffusion Version 2. The Stable Diffusion project, developed by Stability AI, is a cutting-edge image synthesis model that utilizes latent diffusion techniques for high-resolution image generation. It offers an advanced method of generating images based on text input, making it highly flexible for various creative applications. The repository contains pretrained models, various checkpoints, and tools to facilitate image generation tasks, such as fine-tuning and modifying the models. Stability AI's approach to image synthesis has contributed to creating detailed, scalable images while maintaining efficiency. -
16
Stable Diffusion
A latent text-to-image diffusion model
Stable Diffusion is a widely used open-source latent text-to-image diffusion model developed by the CompVis group for generating high-quality images from natural language prompts. The model operates by conditioning a diffusion process on text embeddings produced by a CLIP text encoder, enabling detailed and controllable image synthesis. It was trained on large-scale image datasets and later fine-tuned to produce 512×512 images with strong visual fidelity. Because the system runs efficiently on consumer hardware compared to earlier generative models, it helped popularize local AI image generation workflows. The repository includes reference scripts and model configurations that allow researchers and developers to reproduce, modify, or extend the architecture. Overall, stable-diffusion has become a foundational tool in the generative AI ecosystem for art creation, research, and multimodal experimentation. -
17
Kimi K2
Kimi K2 is the large language model series developed by Moonshot AI
Kimi K2 is Moonshot AI’s advanced open-source large language model built on a scalable Mixture-of-Experts (MoE) architecture that combines a trillion total parameters with a subset of ~32 billion active parameters to deliver powerful and efficient performance on diverse tasks. It was trained on an enormous corpus of over 15.5 trillion tokens to push frontier capabilities in coding, reasoning, and general agentic tasks while addressing training stability through novel optimizer and architecture design strategies. The model family includes variants like a foundational base model that researchers can fine-tune for specific use cases and an instruct-optimized variant primed for general-purpose chat and agent-style interactions, offering flexibility for both experimentation and deployment. With its high-dimensional attention mechanisms and expert routing, Kimi-K2 excels across benchmarks in live coding, math reasoning, and problem solving. -
18
Hunyuan3D-2.1
From Images to High-Fidelity 3D Assets
Hunyuan3D-2.1 is Tencent Hunyuan’s advanced 3D asset generation system that produces high-fidelity 3D models with Physically Based Rendering (PBR) textures. It is fully open-source with released model weights, training, and inference code. It improves on prior versions by using a PBR texture pipeline (enabling realistic material effects like reflections and subsurface scattering) and allowing community fine-tuning and extension. It supports both shape generation (mesh geometry) and texture generation modules. Physically Based Rendering texture synthesis to model realistic material effects, including reflections, subsurface scattering, etc. Cross-platform support (MacOS, Windows, Linux) via Python / PyTorch, including diffusers-style APIs. -
19
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. -
20
DINOv3
Reference PyTorch implementation and models for DINOv3
DINOv3 is the third-generation iteration of Meta’s self-supervised visual representation learning framework, building upon the ideas from DINO and DINOv2. It continues the paradigm of learning strong image representations without labels using teacher–student distillation, but introduces a simplified and more scalable training recipe that performs well across datasets and architectures. DINOv3 removes the need for complex augmentations or momentum encoders, streamlining the pipeline while maintaining or improving feature quality. The model supports multiple backbone architectures, including Vision Transformers (ViT), and can handle larger image resolutions with improved stability during training. The learned embeddings generalize robustly across tasks like classification, retrieval, and segmentation without fine-tuning, showing state-of-the-art transfer performance among self-supervised models. -
21
DeepSeek-V3.2-Exp
An experimental version of DeepSeek model
DeepSeek-V3.2-Exp is an experimental release of the DeepSeek model family, intended as a stepping stone toward the next generation architecture. The key innovation in this version is DeepSeek Sparse Attention (DSA), a sparse attention mechanism that aims to optimize training and inference efficiency in long-context settings without degrading output quality. According to the authors, they aligned the training setup of V3.2-Exp with V3.1-Terminus so that benchmark results remain largely comparable, even though the internal attention mechanism changes. In public evaluations across a variety of reasoning, code, and question-answering benchmarks (e.g. MMLU, LiveCodeBench, AIME, Codeforces, etc.), V3.2-Exp shows performance very close to or in some cases matching that of V3.1-Terminus. The repository includes tools and kernels to support the new sparse architecture—for instance, CUDA kernels, logit indexers, and open-source modules like FlashMLA and DeepGEMM are invoked for performance. -
22
HeartMuLa
A Family of Open Sourced Music Foundation Models
HeartMuLa is the open-source library and reference implementation for the HeartMuLa family of music foundation models, designed to support both music generation and music-related understanding tasks in a cohesive stack. At the center is HeartMuLa, a music language model that generates music conditioned on inputs like lyrics and tags, with multilingual support that broadens the range of lyric-driven use cases. The project also includes HeartCodec, a music codec optimized for high reconstruction fidelity, enabling efficient tokenization and reconstruction workflows that are critical for training and generation pipelines. For text extraction from audio, it provides HeartTranscriptor, a Whisper-based model tuned specifically for lyrics transcription, which helps bridge generated or recorded audio back into structured text. It also introduces HeartCLAP, which aligns audio and text into a shared embedding space. -
23
HunyuanWorld-Voyager
RGBD video generation model conditioned on camera input
HunyuanWorld-Voyager is a next-generation video diffusion framework developed by Tencent-Hunyuan for generating world-consistent 3D scene videos from a single input image. By leveraging user-defined camera paths, it enables immersive scene exploration and supports controllable video synthesis with high realism. The system jointly produces aligned RGB and depth video sequences, making it directly applicable to 3D reconstruction tasks. At its core, Voyager integrates a world-consistent video diffusion model with an efficient long-range world exploration engine powered by auto-regressive inference. To support training, the team built a scalable data engine that automatically curates large video datasets with camera pose estimation and metric depth prediction. As a result, Voyager delivers state-of-the-art performance on world exploration benchmarks while maintaining photometric, style, and 3D consistency. -
24
MiMo-V2-Flash
MiMo-V2-Flash: Efficient Reasoning, Coding, and Agentic Foundation
MiMo-V2-Flash is a large Mixture-of-Experts language model designed to deliver strong reasoning, coding, and agentic-task performance while keeping inference fast and cost-efficient. It uses an MoE setup where a very large total parameter count is available, but only a smaller subset is activated per token, which helps balance capability with runtime efficiency. The project positions the model for workflows that require tool use, multi-step planning, and higher throughput, rather than only single-turn chat. Architecturally, it highlights attention and prediction choices aimed at accelerating generation while preserving instruction-following quality in complex prompts. The repository typically serves as a launch point for running the model, understanding its intended use cases, and reproducing or extending its evaluation on reasoning and agent-style tasks. In short, MiMo-V2-Flash targets the “high-speed, high-competence” lane for modern LLM applications. -
25
HY-World 1.5
A Systematic Framework for Interactive World Modeling
HY-WorldPlay is a Hunyuan AI project focusing on immersive multimodal content generation and interaction within virtual worlds or simulated environments. It aims to empower AI agents with the capability to both understand and generate multimedia content — including text, audio, image, and potentially 3D or game-world elements — enabling lifelike dialogue, environmental interpretations, and responsive world behavior. The platform targets use cases in digital entertainment, game worlds, training simulators, and interactive storytelling, where AI agents need to adapt to real-time user inputs and changes in environment state. It blends advanced reasoning with multimodal synthesis, enabling agents to describe scenes, generate context-appropriate responses, and contribute to narrative or gameplay flows. The underlying framework typically supports large-context state tracking across extended interactions, blending temporal and spatial multimodal signals.
