Build your local AI rig

Firecrawl is the industry-standard tool for turning entire websites into clean, LLM-ready markdown. It handles all the 'messy' parts of web scraping, including JavaScript rendering, proxy rotation, and anti-bot bypass, automatically. Designed specifically for AI developers, it can crawl entire domains and output structured data that is perfectly formatted for RAG (Retrieval-Augmented Generation) or fine-tuning. It acts as the bridge between the unstructured web and the structured needs of modern AI agents.

Crawl4AI is an open-source, high-performance web crawling and scraping engine specifically optimized for large language models. It provides a robust, asynchronous architecture that can handle complex JavaScript-heavy websites, dynamic content, and multi-page crawls with ease. Unlike traditional scrapers, Crawl4AI focuses on 'semantic extraction', automatically identifying the core content of a page and converting it into structured markdown or JSON that is ready for RAG pipelines. It is designed to be deeply integrated into Python-based AI workflows, offering native support for Playwright and advanced proxy management.

ComfyUI is an open-source node-based interface for diffusion models. Instead of a prompt box, a generation is a graph: loaders, samplers, conditioning, upscalers and masks wired together, each step inspectable and re-runnable. Workflows save as JSON and can be shared, which is why most published Stable Diffusion and FLUX pipelines circulate as ComfyUI graphs.

Kimi K3 is Moonshot AI's open-weight model released July 16, 2026, built at roughly 2.8 trillion parameters. It's available now via kimi.com, Kimi Work, Kimi Code, and the Kimi API, priced at $0.30 per million input tokens (cached), $3.00 per million (uncached), and $15.00 per million output tokens. Moonshot published the full open weights on Hugging Face on July 27, 2026, as promised at launch. Moonshot positions it as approaching Anthropic Fable 5-tier performance at a fraction of the cost, though the company acknowledges a tendency toward 'excessive proactivity' on long-running tasks.

Command-line AI coding assistant developed by Anthropic, designed for agentic coding workflows. Unlike traditional IDEs, operates entirely within the terminal, allowing developers to delegate coding tasks directly to Claude AI model. Integrates seamlessly with existing code editors, providing a streamlined coding experience. Enables code generation, debugging, and architectural guidance through natural language commands. Requires Claude Pro or Max subscription and is designed for users comfortable with command-line interfaces. Provides direct interaction with Claude for coding tasks without GUI overhead.

Invoke is an open-source generative image platform combining a unified canvas with inpainting, outpainting and layer control, plus a node editor for building repeatable workflows. It can be self-hosted or used through a commercial hosted tier aimed at studios.

DeepSeek is the architect of the 'DeepSeek movement,' a fundamental shift in AI development that prioritizes extreme efficiency over raw compute. Founded by High-Flyer Quant, they proved that architectural innovations like Multi-head Latent Attention (MLA) and DeepSeekMoE could match the performance of $100B models like GPT-4o and Claude 3.5 while costing 95% less to train and run. Their ecosystem includes the flagship DeepSeek-V3, the reasoning-heavy DeepSeek-R1, and the state-of-the-art DeepSeek-VL2 for high-fidelity OCR and vision tasks. DeepSeek is committed to the open-source community, regularly releasing model weights and technical papers that have democratized frontier-level AI for developers globally.

Llama is Meta AI's open-source large language model family with multiple versions: Llama (February 2023), Llama 2 (July 2023), Llama 3 (April 2024), Llama 3.1 405B (405B parameters, July 2024), Llama 3.3 (December 2024), Llama 4 Maverick (April 2026), and Llama 4 Scout (April 2026). Designed for research and commercial use with strong performance across text generation, reasoning, and code tasks. Available in various sizes from 7B to 405B parameters. Supports multiple languages and extended context windows. Available through Meta's official channels, Hugging Face, and various cloud providers. Open-source licensing allows for local deployment and customization.

Mistral AI provides high-performance large language models with both open-source and commercial offerings. Models include Mistral 7B, Mistral 8x7B (Mixtral), Mistral Large, Mistral Large 2.1, Mistral Small, Pixtral (multimodal, 123B parameters), and the Magistral family (June 2026) - reasoning models designed for enhanced accuracy through increased computational power during inference. Designed for efficiency and performance with strong multilingual capabilities, particularly for European languages. Offers both open-source models for local deployment and commercial API access. Available through Mistral AI's platform, Hugging Face, and various cloud providers. Strong focus on European data privacy and compliance.

MiniMax H3, released 31 July 2026, reads text, images, video and audio as one shared context and returns video with native stereo sound at up to 2K and 15 seconds. MiniMax positions it as omni-modal rather than a video model with features bolted on: text-to-video, image-to-video, editing, reference-based and audio-driven generation are all expressed as natural-language instructions over that one context instead of separate expert models. A single generation accepts up to 9 reference images, 3 reference video clips and 3 reference audio clips, 12 files in total. It runs on the MiniMax platform API as model ID MiniMax-H3 and in the consumer Hailuo app.

Qwen is Alibaba Cloud's family of large language models with multiple versions: Qwen-1.5 (February 2024), Qwen2 (2024), Qwen2.5 (January 3, 2026) with seven dense models from 0.5B to 72B parameters plus MoE variants, and Qwen3 (April 29, 2026) with variants Qwen3-Next, Qwen3-Max, and Qwen3-Omni focusing on context length scaling and parameter efficiency. Designed for multilingual applications with strong support for Chinese, English, and other languages. Excels at code generation, mathematical problem-solving, and structured data understanding. Pre-trained on significantly larger datasets than predecessors. Available through Alibaba Cloud API (DashScope), Hugging Face, and open-source model weights for local deployment. Offers both commercial API access and open-source licensing.

Microsoft Phi is a family of small, efficient language models designed for high performance with minimal parameters. Available models include Phi-1, Phi-2 (December 2023, 2.7B parameters), Phi-3 (April 2024), Phi-3.5, and Phi-4 (2026, 14B parameters) with variants: Phi-4-base, Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini. Marketed as 'small language models' specializing in complex reasoning tasks. Optimized for reasoning tasks, code generation, and efficient inference. Released under MIT license for unrestricted use and modification. Available through Azure OpenAI Service, Hugging Face, and open-source model weights. Designed for edge devices, mobile applications, and cost-effective deployments.

SERA is a family of open-source coding agents developed by the Allen Institute for AI (AI2). It allows developers to customize and fine-tune models on private codebases without exposing sensitive data to external servers. SERA uses synthetic training data to achieve performance comparable to much larger proprietary models at a significantly lower cost.

Gemma is Google DeepMind's family of open-source large language models, serving as lightweight versions of Gemini. Available models include Gemma 1 (February 2024), Gemma 2 (June 2024), and Gemma 3 (March 2026) with variants like PaliGemma for vision-language tasks and MedGemma for medical applications. Available in multiple sizes (2B, 7B, and larger variants). Designed for research, education, and commercial applications with permissive licensing. Trained on similar data and methods as Gemini models but optimized for open-source deployment. Available through Hugging Face, Kaggle, and Google Cloud Vertex AI.

Qwen 2.5-VL is Alibaba's state-of-the-art open-weight multimodal model, designed to bridge the gap between open source and proprietary vision-language models. It features advanced 'NaViVi' (Native Dynamic Resolution) architecture, allowing it to process images of any resolution and videos of any length with extreme precision. It excels at complex visual reasoning, document understanding (OCR), and real-time video analysis, matching or exceeding GPT-4o in many multimodal benchmarks while remaining fully open for the community to build upon.

DBRX is a mixture-of-experts transformer model developed by Databricks and Mosaic ML. Released on March 27, 2024, with 132 billion total parameters (36B active parameters per token). Available in base and instruction-tuned (dbrx-instruct) variants. Outperforms other open-source models in various benchmarks including language understanding, programming, and mathematics. Uses fine-grained mixture-of-experts (MoE) architecture with 16 experts and 4 active per token for efficient inference. Trained at approximately $10 million cost. Released under Databricks Open Model License (permissive for research and commercial use). Available through Databricks Foundation Models API, Hugging Face, and open-source model weights.

Llama 3.2 Vision is Meta's first open-weight multimodal model family, bringing high-fidelity image reasoning to the Llama ecosystem. It integrates vision and text into a unified transformer architecture, enabling it to understand images, charts, and diagrams with the same ease as text. Available in 11B and 90B versions, it is designed for efficiency and edge deployment, making it the industry standard for developers building open multimodal applications that require deep reasoning and broad community support.

Pixtral Large is Mistral AI's flagship 124B parameter multimodal model, designed to compete directly with GPT-4o and Claude 3.5 Sonnet. Built on the Mistral Large 2 foundation, it features a native vision encoder that allows it to reason across text and images with extreme precision. It excels at complex diagram understanding, mathematical reasoning with visual context, and high-fidelity image captioning. Pixtral Large is released under the Mistral Research License, allowing developers to explore frontier-level vision-language capabilities with open weights.

InternVL 2.5 is a world-class open-source multimodal large language model (MLLM) that consistently tops the leaderboards for open-weight vision reasoning. It features a powerful 78B parameter architecture with a specialized vision-language alignment that excels at OCR, document understanding, and complex visual Q&A. It is designed to bridge the gap between open models and GPT-4V, offering exceptional performance across a wide range of multimodal benchmarks while remaining fully open for community development.

OpenCode is an open-source terminal-based coding agent that connects to multiple language models and performs autonomous software engineering tasks from the command line. It edits files, runs tests, manages git workflows, and iterates on code with minimal human intervention. The v1.17.8 release from June 2026 improves tool calling reliability, multi-file refactoring, and support for local and remote model backends.

DeepSeek V4-Pro is a high-performance language model from DeepSeek. The model itself was released on April 24, 2026, and permanent pricing was announced on May 31, 2026. It offers strong reasoning and coding performance at a competitive price point, with open weights available for local deployment.

MiniMax M3 is a 1-million-token-context agentic frontier model released on May 31, 2026. It supports long-document analysis, coding, multi-turn agent workflows, and tool use, positioning it as a general-purpose assistant with an exceptionally large context window.

OpenShell is NVIDIA's open-source runtime for running AI agents inside a sandbox. It sits beneath the agent harness, runs the agent without privileges, limits which files it can reach, filters system calls at the kernel level, and checks every network request against policy by binary, destination, method and path. Nothing is allowed by default, credentials are only supplied where policy permits, a policy-aware router controls which model endpoints the agent can call, and every decision is logged. It supports Claude Code, Codex, GitHub Copilot CLI, Hermes, LangChain Deep Agents, OpenClaw and OpenCode, and runs on Docker, Podman, Kubernetes via Helm, and NVIDIA DGX Spark and DGX Station. First announced in March 2026, it became the software half of NVIDIA's Open Agent Safety Platform on...

Generates high-quality images from text prompts using Alibaba's Tongyi Qianwen 20-billion parameter MMDiT model. Excels at complex text rendering with commercial-grade quality, supporting multi-line layouts and paragraph-level text generation in both Chinese and English. Provides advanced image editing capabilities including style transfer, object insertion/removal, and detail enhancement. Ranks first in multiple public benchmark tests, surpassing similar open-source models with superior prompt understanding and visual quality.

FLUX.2 Pro is the definitive answer to closed-source image generators like Midjourney. Developed by Black Forest Labs (the original creators of Stable Diffusion), it represents the pinnacle of high-fidelity, open-weight image generation. It features a massive 12B parameter 'Flow' architecture that produces photorealistic textures, perfect human anatomy, and industry-leading text rendering. Unlike its competitors, FLUX is built for the open-source community, supporting LoRA training, ControlNet, and local deployment, allowing creators to maintain full control over their artistic style and data.

Alibaba's latest open-weight model specialized for coding. At 80B parameters, it matches proprietary performance for local development and autonomous coding agents.

An open-source research stack spanning Python, JS, C++, and CUDA, engineered from the ground up by autonomous AI coding agents. Optimized for high-performance tensor operations.

SAM3D v2 leverages Meta's latest Segment Anything technology to reconstruct 3D geometry from single or multiple images with extreme precision. It is the industry standard for research-grade 3D reconstruction.

Qwen 3 is a 2025 open-weight Mixture-of-Experts model family from Alibaba Cloud, ranging from 0.6B to 235B parameters. It supports both thinking and non-thinking modes, strong multilingual performance, and agentic tool use, and is released under permissive licenses.

Qwen 2.5-Coder is a code-focused open-weight model from Alibaba, available in sizes from 1.5B to 32B parameters. It is optimized for code generation, completion, and debugging across many programming languages and is competitive with closed coding models.

Qwen-Math is a family of open-weight models specialized for mathematical reasoning and problem solving, derived from Qwen 2.5 and optimized on math datasets. It is available in several sizes and is competitive on math benchmarks.

Wan 2.0 is an earlier open-source video generation model from Alibaba, predecessor to Wan 2.1. It supports text-to-video and image-to-video generation and laid the groundwork for the Wan family's open-source release.

FLUX.2 [schnell] is the fastest open-weights FLUX.2 variant, designed for 4-8 step local inference on consumer hardware. It retains strong prompt adherence and text rendering while being freely available for local and commercial use.

FLUX.2 [dev] is an open-weight FLUX.2 model from Black Forest Labs, released for non-commercial and commercial research. It offers a strong balance of quality and efficiency, making it the base for many fine-tunes and community LoRAs.

BRIA FIBO is an 8B-parameter DiT text-to-image model trained on long structured JSON captions. It turns short prompts into detailed structured schemas and generates images with precise, reproducible control over composition, lighting, camera, and color. It is also available in an image-to-image 'Inspire' mode and is trained entirely on licensed data for commercial safety.

BRIA FIBO Lite is a lightweight variant of the FIBO image generation pipeline. It pairs an open-source FIBO-VLM bridge with a smaller FIBO Lite model to enable rapid inference and fully local, on-prem deployment for privacy-critical environments.

BRIA RMBG 2.0 is a dichotomous image segmentation model that produces a grayscale alpha matte for high-quality background removal. It is trained on over 15,000 fully licensed, manually labeled high-resolution images and is designed for e-commerce, advertising, gaming, and enterprise content workflows.

DeepSeek V4-Flash is the efficient sibling of V4-Pro, offering a 1M-token context window and configurable thinking modes at a fraction of the API cost. It is optimized for high-throughput chat, classification, bulk extraction, and agentic coding workloads, with a July 2026 update that boosted agent and coding benchmarks.

DeepSeek R1 is a 671B-parameter open-weight reasoning model that matches o1-class performance on math, code, and logic benchmarks through reinforcement learning on verifiable tasks. It exposes chain-of-thought reasoning and is available as MIT-licensed local weights and via API, with the R1-0528 update in May 2025 further improving math and code reasoning.

DeepSeek V3.2 is a 128K-context mixture-of-experts model that unified thinking and non-thinking modes in the V3 line and introduced DeepSeek Sparse Attention. It served as the December 2025 flagship before V4 and remains available for self-hosting and as a historical comparison point.

GLM-5 is Zhipu AI's (Z.ai) first 2026 flagship, a 744B-parameter sparse mixture-of-experts model with roughly 40B active parameters per token. It is built for high-intelligence reasoning, agentic planning, and long-context execution, with a 200K context window and 128K maximum output.

GLM-5.1 is Z.ai's refinement flagship released in April 2026, a 744B-parameter MoE model with 40B active parameters per token. It targets long-horizon agentic coding, multi-file refactoring, and terminal work, sustaining up to 8-hour autonomous tasks through a 200K context window and 128K maximum output.

Gemma 3 is an open-weights family of multimodal models from Google, ranging from 1B to 27B parameters. It supports text and image input, a 128K context window, and is released under a permissive license for research and commercial use.

Open-source Transformer-based Mixture-of-Experts language model with 389 billion total parameters and 52 billion active parameters. Supports instruction-tuned and long-context pretraining checkpoints up to 256K tokens, distributed via Hugging Face and GitHub.

Hybrid Mamba-Transformer MoE reasoning model released March 2025, built on Hunyuan TurboS with 52 billion active parameters and a 256K context window. It focuses compute on reinforcement-learning post-training and scores strongly on math, coding, and graduate-level reasoning tasks.

A 200K-context open-weight Hunyuan model optimized for speed while maintaining strong performance on general chat, coding, and agentic tasks. It also serves as the base for the Hunyuan T1 reasoning model.

Open-weight instruction-tuned variant of Tencent's Hunyuan 2.0 series, offering a 131K-token context window and strong everyday performance for chat, content creation, coding, and enterprise workflows.

Open-weight reasoning variant of Hunyuan 2.0 with a 131K context window, designed for complex problem-solving, math, and long-context reasoning workflows.

A compact open-weight MoE instruction model with a 131K context window, listed as a cost-efficient everyday Hunyuan option via OpenRouter and Tencent Cloud.

Open-weight preview of Hunyuan 3 (Hy3), a 295B-parameter MoE model with 21B active parameters and a 262K context window. Supports reasoning, function calling, and tool use, and is available via OpenRouter and Tencent Cloud TI Platform.

Open-source text-to-image diffusion transformer with fine-grained Chinese and English understanding, multi-turn prompt refinement, ControlNet, LoRA, and IP-Adapter support. Available via Hugging Face, Diffusers, ComfyUI, and a web demo.

Native multimodal MoE image generation model with 80 billion total parameters and 13 billion active parameters. It unifies multimodal understanding and generation within an autoregressive framework, supports image editing and multi-image fusion, and is released as the largest open-source image generation model.

Generates explorable, interactive 3D worlds from text prompts or images using panoramic proxies and semantic layering. Supports mesh export for game engines, VR/AR, and interactive content creation.

Open-source system for generating high-resolution textured 3D assets from text or images. Includes a shape-generation diffusion transformer, texture synthesis pipeline, PBR support, and a Blender add-on, with mini and multiview variants for different hardware.

Kimi K2.5 is Moonshot AI's open-weight multimodal model released January 27, 2026. It supports text, image, and video input, thinking and non-thinking modes, and a 256K context window, with strong performance on agent, coding, and vision tasks. Moonshot announced the kimi-k2.5 API will be retired on August 31, 2026, so production workloads should plan a migration path to K2.6 or K3.

Kimi K2.6 is Moonshot AI's open-weight multimodal model released April 21, 2026. It supports text, image, and video input, thinking and non-thinking modes, and a 256K context window, with improved long-context coding stability and agent-task performance.

Kimi K2.7 Code Highspeed is the high-speed serving variant of Moonshot AI's K2.7-Code model, released June 12, 2026. It delivers roughly 180 tokens per second for coding tasks while preserving the same 256K context window, text/image/video input, and thinking-mode capabilities.

Llama 4 Maverick is Meta's flagship open-weight model, released in April 2025 as part of the Llama 4 family. It uses a Mixture-of-Experts architecture with 17B active parameters and around 400B total parameters, natively understands text and images, and supports a 1M-token context window.

Llama 4 Scout is Meta's efficient Llama 4 variant, released in April 2025. It is a Mixture-of-Experts model with 17B active parameters and 109B total parameters across 16 experts, natively multimodal for text and images, and supports an industry-leading 10M-token context window.

Llama 3.3 is Meta's 70B-parameter multilingual instruction-tuned model, released in December 2024. It delivers text-only performance comparable to the much larger Llama 3.1 405B model while running far more efficiently, with a 128K-token context window and broad language support.

Llama 3.1 405B is Meta's 405-billion-parameter dense open-weight model, released in July 2024. It was the first openly available model to reach frontier-level performance on reasoning, coding, and multilingual tasks, with a 128K-token context window and native tool-use support.

MiniMax M2.7 is a general-purpose language model built for real-world engineering, professional office tasks, and character-rich interaction. It is positioned as MiniMax's mid-tier coding and agentic model alongside the larger M3.

A 675B-parameter sparse mixture-of-experts model with 41B active parameters and a 262K context window, released under Apache 2.0. It handles text and vision tasks, supports strong multilingual performance, and is designed for both research and enterprise deployment.

A 119B-parameter MoE model with 6B active parameters and a 256K context window, released under Apache 2.0. It unifies instruct, reasoning, multimodal, and agentic coding capabilities in a single efficient model with configurable reasoning effort.

A code generation model optimized for latency-sensitive fill-in-the-middle completion and chat, supporting 80+ programming languages. It is designed for IDE integration and enterprise software development workflows.

A family of 3B, 8B, and 14B parameter dense models released under Apache 2.0, optimized for performance-to-cost ratio at the edge. The 14B variant includes reasoning capabilities, making the family suitable for on-device and local deployments.

A family of open-weight speech models including a 24B production variant and a 3B edge variant, released under Apache 2.0. It supports transcription, audio understanding, summarization, Q&A, and function calling from voice with multilingual support.

A 30B total / 3B active parameter hybrid Mamba-2 + Transformer MoE language model built for efficient on-device and edge agentic tasks. It features a 1M-token context window, reasoning ON/OFF modes with configurable thinking budgets, and up to 4× faster throughput than its predecessor.

A 120B total / 12B active parameter hybrid Mamba-Transformer MoE language model with LatentMoE, multi-token prediction, and native NVFP4 pretraining. Optimized for complex multi-agent applications with a 1M-token context window and up to 5× higher throughput than the previous Nemotron Super.

An 8B-class physical-AI omni-model (8B reasoner + 8B generator) optimized for efficient inference on workstation-grade NVIDIA hardware such as the RTX PRO 6000. It unifies vision reasoning, world generation, and action prediction for robotics and physical AI prototyping.

A 4B-class physical-AI omni-model (2B reasoner + 2B generator) optimized for real-time robotic policy and visual reasoning at the edge. It unifies world generation, vision reasoning, and action prediction in a compact form factor for embedded deployment.

A general-purpose document parsing model that overcomes traditional OCR limitations by understanding complex page layouts. It extracts structured text, tables as Markdown or LaTeX, bounding boxes, and semantic classes from unstructured PDFs and images.

A 4B-parameter multimodal, multilingual small language model designed as a robust content-safety moderator. It supports standard taxonomy safety classification and custom-policy enforcement with reasoning traces for text and image inputs.

A 253B-parameter variant of Llama 3.1 fine-tuned by NVIDIA using the HelpSteer2 datasets to improve helpfulness and instruction adherence. It is the largest member of the Llama-3.1-Nemotron family of community collaboration models.

A 49B-parameter variant of Llama 3.1 fine-tuned by NVIDIA using the HelpSteer2 datasets to improve helpfulness and instruction adherence. It is the mid-size member of the Llama-3.1-Nemotron family, optimized for a strong performance-to-size ratio.

An 8B-parameter variant of Llama 3.1 fine-tuned by NVIDIA using the HelpSteer2 datasets to improve helpfulness and instruction adherence. It is the smallest member of the Llama-3.1-Nemotron family, optimized for efficient on-device and edge deployment.

NVIDIA's open foundation model for humanoid robot reasoning and control, combining an Eagle-based vision-language backbone with a diffusion transformer (DiT) action head for language-conditioned manipulation across diverse embodiments.

Stable Diffusion 1.5 is the landmark open-source latent diffusion model released by Stability AI in 2022. It established the open image generation ecosystem and remains the base for countless fine-tunes, LoRAs, and ControlNet models.

Stable Diffusion XL (SDXL) is a 2023 open-source text-to-image model that generates higher-quality, higher-resolution images than SD 1.5. It uses a two-stage base-plus-refiner pipeline and is widely used for production image workflows.

Stable Diffusion XL Turbo is a distilled, fast variant of SDXL that can generate images in a single step or a few steps. It is optimized for low-latency applications and real-time interactive generation.

Stable Diffusion 3 is a 2024 text-to-image model from Stability AI based on a Multimodal Diffusion Transformer architecture. It improved text rendering, composition, and prompt adherence over SDXL.

Stable Diffusion 3 Medium is a 2B-parameter version of SD3 designed to run well on consumer GPUs. It offers a balance of SD3 quality and efficiency, making it accessible for local creators.

Stable Diffusion 3.5 Large is an 8B-parameter open-weights model in the SD 3.5 family, offering the highest quality and best prompt adherence of the 3.5 series for demanding image generation tasks.

TRELLIS Mini is a smaller, faster variant of the TRELLIS family from Microsoft Research, designed for efficient image-to-3D generation on limited hardware while preserving the core architecture's quality.

TRELLIS Large is the larger, higher-quality variant of the TRELLIS family, producing detailed 3D assets from single images with better geometry and texture fidelity than the smaller variants.

TripoSR is an open-source, feed-forward image-to-3D model developed by Stability AI and Tripo AI. It generates textured 3D meshes from a single image in under a second on a single GPU and is released under an MIT license.

NanoClaw is an open-source personal AI agent runtime built by Gavriel Cohen as a smaller, auditable alternative to OpenClaw. Each agent session runs in its own Docker container with scoped permissions and self-destructs when the task ends. In August 2026 it added a Slack integration that lets you provision persistent AI agent teams and colleagues from a single message, running on customer infrastructure.

TRELLIS 2 is an open-source image-to-3D generation model from Microsoft Research, released in 2026. It reconstructs 3D assets from single images or text prompts and is designed for research and experimentation.

NVIDIA Cosmos 3 is an open physical-AI omnimodel released around May 31 to June 1, 2026. It generates video, 3D, and physical-world simulations to train and evaluate robotics and autonomous vehicle systems without expensive real-world data collection.

Microsoft TRELLIS generates high-quality 3D models from text prompts or reference images using a unified Structured LATent (SLAT) representation. Trained on 500,000 3D objects, it produces detailed 3D assets in multiple formats including meshes, radiance fields, and 3D Gaussians. Supports flexible editing capabilities for generating variants and localized modifications. Production-ready outputs with proper topology, UV mapping, and textures suitable for game engines, VR applications, and digital content creation. Integrated with NVIDIA AI Blueprint for accelerated 3D generation.

LTX-2.5 is an open-weights video and world model from LTX (spun out of Lightricks). Generates 10-second video clips from images in 6.8 seconds on Nvidia superchips. Features multi-shot support, diffusion decoder for higher quality, new conditioning modes, and autoregressive models for real-time use and robotics. Weights available on Hugging Face and through LTX API. Free for organizations under $10M ARR; larger companies negotiate license. Over 33 million downloads of LTX models—most-used open world model line.

Atria Dawn Preview is an open-weight preview release from Shanghai AI Laboratory, released around September 11, 2026 under the MIT license. Built on the 744B-parameter MoE GLM-5.2 foundation model, it accepts text input only with a 256K-token context window and is purpose-built for long-horizon research and engineering agents: combining task objectives with environmental feedback to carry a scientific method through problem analysis, tool use, code implementation, experiment execution, result analysis and failure recovery. Across 16 published benchmarks spanning real-world research, engineering and digital work, Shanghai AI Lab reports it is competitive with frontier agents and achieves the top reported score on five of them.

Generates and edits images via an open model ecosystem including Stable Diffusion models and community tools. Provides local generation, API access, and extensive customization options with fine control over generation parameters. Supports multiple model versions, LoRA fine-tuning, ControlNet for precise control, and a vast ecosystem of community models and tools. Enables complete workflow customization from local deployment to cloud API integration, making it the foundation for many custom image generation pipelines.

Alibaba released a 7-billion parameter language model optimized for consumer laptop deployment, released August 2026. Quantized to 4-bit with custom ONNX optimization for CPU/GPU inference. Competitive response to Meta's on-device model strategy, targeting Windows/Mac laptops with 8GB+ RAM. Open weights under OpenMDW-1.1 license. Achieves reasonable performance on everyday tasks (email drafting, code generation) while running entirely offline without cloud dependency.

LTX-2.5 is an open-weights video and world model from Lightricks (LTX company spun out of Lightricks), released August 2026. Generates 10-second video clips from images in 6.8 seconds on Nvidia superchips. Features multi-shot support, diffusion decoder for higher quality, new conditioning modes, and autoregressive models for real-time use and robotics. Weights freely available on Hugging Face under OpenMDW-1.1 license. 33 million downloads, most-used open world model line on the market. Free for organizations under $10 million annual revenue; larger companies negotiate licenses.

SWE-2 is Cognition's flagship coding model, released September 12, 2026 and built by reinforcement-learning post-training on top of Moonshot AI's 2.8-trillion-parameter open-weight Kimi K3. Cognition reports it matches Claude Fable 5.1 on FrontierCode 1.1 Main (both around 50%) at roughly 64% lower inference cost, scores 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, but trails Fable 5.1 and GPT-6 Astra by around 30 points on the harder Terminal-Bench 4. It ships with no open weights and no self-serve public API: it is exclusive to Devin's desktop app and CLI, with a Devin Web and Fusion rollout planned. Usage inside Devin is free through mid-October 2026; after that it requires a paid Devin plan, and separate enterprise credit-based API access is priced at $3.00 per million input...

Google Gemma is a family of open-source language models available in 2B, 7B, and 27B parameter sizes. Trained on 6 trillion tokens of high-quality data, designed for efficient local deployment. Includes standard and instruction-tuned variants. Reached 1 billion total downloads across Hugging Face, Kaggle, and GitHub by August 2026. Has spawned 100,000+ community derivatives (finetuned versions, specialized variants, quantizations). Available under Google DeepMind's permissive license for commercial and research use.

Generates 3D objects from text prompts or images using OpenAI's Shap-E model, a conditional generative model for 3D assets. Produces high-quality 3D meshes, point clouds, and neural radiance fields (NeRFs) from natural language descriptions. Supports both text-to-3D and image-to-3D workflows, generating detailed 3D models with realistic geometry and textures suitable for game assets, product visualization, and 3D printing applications. Open-source model with comprehensive documentation and active community support, making it ideal for research, prototyping, and educational use.

Muse Glimmer is Meta's 30 billion parameter open-weight model released August 2026, designed to run autonomous AI agents directly on consumer hardware without cloud dependency. It ships under the permissive Apache 2.0 license (Meta's first fully open release since moving to proprietary Muse Spark in April), quantized to roughly 4 bits with block-level speculative decoding for fast local inference. Drops the 700M monthly user restriction that hampered prior Llama releases.

Generates 3D point clouds from text prompts using OpenAI's Point-E model, a fast and efficient approach to 3D generation. Produces detailed point cloud representations of 3D objects from natural language descriptions, enabling rapid iteration and exploration of 3D concepts. Optimized for speed while maintaining quality, making it ideal for quick prototyping, concept exploration, and applications requiring fast 3D asset generation workflows. Efficient architecture enables fast inference times compared to mesh-based generation, making it perfect for early-stage 3D concept exploration.

Generates high-quality 3D NeRF (Neural Radiance Field) representations from text prompts using score distillation sampling, a technique that leverages pre-trained 2D diffusion models for 3D generation. Produces detailed 3D scenes and objects with realistic lighting, materials, and geometry from natural language descriptions. Enables creation of view-consistent 3D content without requiring 3D training data, making it ideal for generating complex 3D scenes, objects, and environments for visualization, games, and virtual reality applications. Pioneering approach uses 2D diffusion models to guide 3D NeRF generation, enabling high-quality 3D creation from text.

Generates high-quality 3D meshes with textures from images or text using NVIDIA's Get3D model, a generative model that produces detailed 3D triangular meshes with high-resolution textures. Creates production-ready 3D assets with proper topology, realistic materials, and fine geometric details suitable for game engines, 3D software, and real-time rendering applications. Supports both image-to-3D and text-to-3D workflows, generating textured meshes that can be directly exported to standard 3D formats. NVIDIA's research-grade model with exceptional quality, making it ideal for production workflows requiring game-ready 3D assets.

Laguna S 2.1 is a 118 billion parameter Mixture-of-Experts model from Poolside AI, released August 2026. Activates only 8 billion parameters per token, supports context windows up to 1 million tokens, runs under permissive OpenMDW-1.1 license. Benchmarks: 70.2% on Terminal-Bench 2.1 (beating DeepSeek-V4-Pro-Max, Nvidia Nemotron 3 Ultra), 78.5% on SWE-Bench Multilingual. Trained in under 9 weeks on 4,096 Nvidia H200 GPUs.

Generates 3D models from single images using Zero-1-to-3, a model that learns to generate novel views of objects from a single input image. Produces view-consistent 3D representations by understanding object geometry and appearance from limited input. Enables creation of 3D assets from photographs, product images, or concept art, making it ideal for 3D reconstruction, product visualization, and asset generation workflows. Advanced geometric understanding enables high-quality 3D reconstruction from single images with view consistency across different angles.

Generates 3D models from single images using Instant3D, a fast and efficient approach to image-to-3D conversion. Produces detailed 3D meshes with textures from photographs in minutes, enabling rapid prototyping and asset creation. Optimized for speed while maintaining quality, making it suitable for quick iterations, concept exploration, and workflows requiring fast 3D asset generation from reference images. Efficient architecture enables rapid 3D mesh creation, making it ideal for workflows prioritizing speed and rapid iteration.

Baidu ERNIE 4.5 (Enhanced Representation through Knowledge Integration) is a family of large language models released by Baidu in November 2026. The ERNIE 4.5 model family includes 10 variants ranging from 0.3 billion to 424 billion total parameters, utilizing a Mixture-of-Experts (MoE) architecture for efficient inference. Open-sourced under the Apache 2.0 license in June 2026, ERNIE 4.5 demonstrates strong performance in Chinese natural language processing, multimodal understanding, and various AI benchmarks. The model excels in common-sense reasoning, optical character recognition, and Chinese language tasks. Available through ERNIE Bot (web interface) and Baidu's Qianfan platform (API access), with open-source model weights available for local deployment.

GLM-4.5 (General Language Model) is Zhipu AI's latest large language model in the ChatGLM/GLM series, released in 2026. Building on the success of previous GLM models, GLM-4.5 offers enhanced reasoning capabilities, improved multilingual support (with strong Chinese and English capabilities), and advanced instruction-following. The model is designed for both chat and completion tasks, with support for long context windows and fine-tuned variants for specific use cases. GLM-4.5 maintains Zhipu AI's focus on efficient inference and cost-effective deployment. Available through Zhipu AI's platform (web interface and API) with options for local deployment of open-source variants.

Hymotion 1.0 (also known as HY-Motion 1.0 or Hunyuan Motion 1.0) is Tencent's open-source, billion-parameter text-to-3D motion generation model released in December 2026. Built on a Diffusion Transformer (DiT) architecture with flow matching, it generates high-fidelity, smooth, and diverse 3D character animations from natural language descriptions. Trained on over 3,000 hours of diverse motion data covering 200+ motion categories including locomotion, sports, fitness, social interactions, and daily activities. The model employs a three-stage training paradigm: large-scale pretraining, high-quality fine-tuning with 400 hours of curated text-motion pairs, and reinforcement learning for physical plausibility. Supports standard 3D formats (FBX, BVH, GLTF) for seamless integration with...

Nemotron 3 Ultra is the largest member of NVIDIA's Nemotron 3 family, released 4 June 2026 at Computex. It is a 550B-parameter hybrid latent mixture of experts with roughly 55B parameters active per token, combining Mamba and Transformer blocks, and trained in NVIDIA's 4-bit NVFP4 format on Blackwell hardware. It ships under the NVIDIA Open Model License, which permits commercial use, and NVIDIA published training data, reinforcement learning environments and post-training recipes alongside the weights rather than the weights alone. Weights are on Hugging Face as nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B in both BF16 and NVFP4, and it is served through OpenRouter, Together AI, Baseten, DeepInfra, Fireworks and NVIDIA NIM.

GLM-5.2 is Zhipu AI's (Z.ai) flagship open-weight model, released 13 June 2026 under an MIT licence with weights published on Hugging Face at zai-org/GLM-5.2. It is a 753B-parameter mixture-of-experts model activating roughly 40B parameters per token, with a 1M-token context window and 128K maximum output. The headline architectural change is IndexShare, which reuses the same indexer across every four sparse attention layers; Z.ai reports this cuts per-token compute by about 2.9x at full 1M context. It targets long-horizon agentic coding, multi-file refactors, terminal work, and tasks that run for many steps rather than single completions.

Inkling is the first open-weights model from Thinking Machines Lab, released 15 July 2026 under Apache 2.0. It is a 975B-parameter mixture of experts with 41B active per token, trained from scratch on 45 trillion tokens of text, images, audio and video, with a context window up to 1M tokens. The architecture uses 256 routed experts plus 2 shared experts per layer with 6 routed experts active per token, a sigmoid router, and interleaved sliding-window and global attention at a 5:1 ratio. It accepts text, image and audio input natively and returns text. Weights are on Hugging Face in both the original format and an NVFP4 checkpoint for Blackwell hardware. A distilled Inkling-Small followed on 31 July.