AI Building Blocks
More actions
| AI Building Blocks | |
|---|---|
| Information | |
| Type | Drag-and-drop Unity components (Building Blocks) for machine-learning features |
| Industry | Virtual reality, mixed reality |
| Developer | Meta Platforms |
| Written In | C# |
| Operating System | Meta Horizon OS (Meta Quest apps built with Unity) |
| License | Meta Platform Technologies SDK License Agreement[1][2] |
| Supported Devices | Meta Quest 3, Meta Quest 3S (Meta's recommendation) |
| Release Date | 10 December 2025 (Meta XR Core SDK 83.0) |
| Website | https://developers.meta.com/vr/documentation/unity/unity-ai-building-blocks-overview/ |
AI Building Blocks are ready-made components that Meta includes in its Meta XR Core SDK for the Unity game engine to add machine-learning features to apps for Meta Quest headsets. Meta's documentation describes four core blocks: object detection in passthrough or camera images, large language models (LLMs) for text and multimodal chat, speech-to-text and text-to-speech.[3][4] The collection first shipped in version 83.0 of the Core SDK, dated 10 December 2025,[5] after Meta had announced an object detection block and an LLM block for that version at Meta Connect 2025.[6]
Each block pairs an agent, a Unity component that handles input, output and inference calls, with a provider, a configuration asset that sets where the model runs: on a cloud service, on a computer in the same local network, or on the headset itself through the Unity Inference Engine (see Unity Sentis).[3][7] Meta supplies providers for the Llama API with its own Llama models and for OpenAI, Hugging Face, ElevenLabs, Roboflow and Ollama.[7] According to the documentation, moving an agent from a cloud model to an on-device model only requires assigning a different provider asset, with no code changes.[4] Meta recommends Unity 6 or newer and a Meta Quest 3 or Meta Quest 3S headset.[3]
The blocks can be combined into pipelines. Meta's documentation gives two examples: a user's speech is transcribed, sent to an LLM and spoken back by the text-to-speech block, or objects detected in a camera frame are passed to an LLM as context for a spoken reply.[4] The vision blocks take their images from the headset's front-facing passthrough cameras, which apps can access through Meta's Passthrough Camera API.[6][4]
History
Announcement
Meta's Building Blocks are pre-built components for common XR features, grouped by category in the Building Blocks dialog of the Unity Editor.[8] In its developer blog post for Meta Connect 2025, dated 18 September 2025, Meta described Building Blocks as a Unity tool, part of the Meta SDK, for adding features such as passthrough or multiplayer "with just a few clicks". The same post announced "four new building blocks coming in v83", two of which it called AI blocks. The Object Detection block was to feed images from the passthrough camera to an object detection algorithm to identify common objects in a room, and the LLM block was to let developers "connect to the LLM of your choice". The other two, Passthrough Camera Access and Passthrough Camera Feed Visualizer, were to give access to a single front-facing camera and show the textures of the left and right cameras.[6]
Meta's recap of Unity's Unite 2025 conference, published on 25 November 2025, says that its developer advocates Jake Steinerman and Dilmer Valecillos gave "a first look at AI Building Blocks" in their session "Unlock the full potential of Meta Quest"; the recap describes the blocks as letting developers bring their own LLM into an app "with a few fields and configuration changes". According to the recap, software engineers Joe Paley and Neel Bedekar introduced the same four new blocks in Meta's final session at the conference, "Accelerating VR Development with Meta Horizon Tools and AI".[9]
Release in version 83
Version 83.0 of the Meta XR Core SDK, dated 10 December 2025, introduced the "AI Building Blocks Collection" for adding "runtime AI features" such as object detection and LLM integration to apps.[5] Meta's AI Building Blocks documentation describes this first release as four core blocks (Object Detection, LLM, Speech to Text and Text to Speech) with a provider system for cloud, local and on-device inference, task interfaces, depth integration, non-maximum suppression accelerated on the GPU and an event-driven architecture.[3] The version of the overview dated 11 December 2025 also stated that cloud and local providers connected through HTTP requests, which Meta attributed to a lack of native WebSocket and WebRTC support in Unity. It warned of increased latency when building real-time experiences with cloud or local inference and said that for this reason Meta would focus more on building out on-device inference with the Unity Inference Engine.[10] The page as updated on 8 July 2026 no longer contains this passage.[3]
ElevenLabs announced on 11 December 2025 that it was partnering with Meta to "power expressive, scalable audio across Instagram, Horizon, and more".[11] On 17 December 2025 the language-industry publication Slator reported that, according to a separate update from Meta's Reality Labs team, ElevenLabs had become a pre-integrated text-to-speech and speech-to-text provider in what Slator called Meta's "AI Inference Building Blocks", a system for connecting external AI models to Quest applications without custom integrations.[12]
Version 85 and later
Version 85.0 of the Core SDK (10 February 2026) added an image segmentation block, which the release notes say provides 2D image segmentation and "support for accurate 3D bounding boxes/segmentations in world space".[13] The AI Building Blocks overview of 8 July 2026 lists further changes under the heading "Version 85 and later": on-device text generation with small language models (SmolLM, Qwen and Phi), central storage of API keys in a CredentialStorage asset, a configurable detection interval for the object detection agent, and image capture events, vision support detection and automatic warm-up of on-device providers in the LLM agent. The providers and agents pages mark the credential storage and the detection interval as available "since v85".[7][4] The chat interface gained request, response and streaming types, and its main method changed in a way that breaks version 83 code: RunChatAsync, which took a prompt string and returned the reply as a string, was replaced by ChatAsync, which takes a ChatRequest, can report streamed partial results and returns a ChatResponse. Task methods received task-specific names such as ChatAsync() and DetectAsync().[3]
At the Game Developers Conference in March 2026, developers could try Building Blocks at Meta's booth. Meta's recap of the second conference day, dated 11 March 2026, said that the new AI Building Blocks let developers plug in a preferred LLM and pair it with Meta's Voice SDK for conversation and voice control.[14] Its closing recap of 13 March 2026 said that Meta's sessions had covered AI Building Blocks for mixed reality apps that rely on object detection and computer vision.[15]
The release notes for version 201.0 (15 April 2026) introduce standardized "OS versioning" (for example "OS 200, 201") and list a fix for the AI Building Blocks that "improved Unity Inference Engine compatibility and setup rules".[16] Version 203.0 (3 June 2026) raised the minimum supported Unity version of the Core SDK to 6000.0.66f2.[17] Meta's recap of Unite Seoul, held on 21 July 2026, lists "AI Building Blocks for integrating on-device AI" and Passthrough Camera Access Building Blocks under the heading "Building Blocks (Now with AI)", among updates to Meta's core Unity tools that it described as shipping in Meta XR SDK v205.[18]
Version history
The table lists the Core SDK releases whose release notes or documentation name changes to the AI Building Blocks.
| Core SDK version | Date | AI Building Blocks changes |
|---|---|---|
| 83.0 | 10 December 2025 | First release of the collection: Object Detection, LLM, Speech to Text and Text to Speech blocks with cloud, local and on-device providers[5][3] |
| 85.0 | 10 February 2026 | Image segmentation block;[13] CredentialStorage and detection interval setting, which the documentation marks "since v85"[7][4] |
| 201.0 | 15 April 2026 | Improved Unity Inference Engine compatibility and setup rules[16] |
Architecture
Meta's documentation separates "what you run (the Agent) from where it runs (the Provider)".[7] An agent is a Unity MonoBehaviour component that captures input, triggers inference and dispatches the results. A provider is a ScriptableObject asset that defines the inference backend and the structure of the input and output, so that agents can stay provider-agnostic and work with any provider that supports the same task.[7][19] Providers implement one or more task interfaces (IChatTask, IObjectDetectionTask, ISpeechToTextTask and ITextToSpeechTask), each exposing a single asynchronous method such as ChatAsync() or DetectAsync(), and all providers in the framework inherit from a base class, AIProviderBase.[19] The code examples in Meta's documentation use the Meta.XR.BuildingBlocks.AIBlocks namespace.[4]
Different agents in one app can use different providers; Meta's example is object detection running locally while a cloud LLM generates text.[20] The documentation's example of changing backends is to prototype with Llama 4 Maverick through the Llama API and later switch to Llama 3.3 running on Ollama, or to an on-device model, without changing the app's logic.[3]
Inference types
Meta's documentation defines three inference types:[7]
| Inference type | Where the model runs | Typical use, according to Meta |
|---|---|---|
| Cloud | A hosted model, reached over HTTPS | Fastest way to prototype with current models, including LLMs, speech models and DETR object detection |
| Local | A model on a computer in the same Wi-Fi network, for example one served by Ollama | Low latency, private demos and exhibitions |
| On-device | The headset itself, through the Unity Inference Engine | Lowest latency, full privacy and no network dependency, with what Meta calls a "significant performance impact" |
Setup
When a developer adds an AI block from the Building Blocks section of Meta Hub, the installer detects which inference types the block supports and loads compatible providers through RemoteProviderProfileRegistry, which retrieves official provider profiles and defaults, such as endpoints and model names, from Meta's content delivery network. The developer chooses cloud, local or on-device inference, and the selected provider asset is saved with the block's prefab or component, where it can later be edited in the Inspector.[7] Provider assets can also be created from Unity's Create menu under Meta, AI, Provider Assets. A cloud provider needs an API key, an endpoint and a model ID. Since version 85, keys can be kept in a shared CredentialStorage asset, with an option to override the key for an individual provider, and provider inspectors can check that a key and endpoint work.[7]
Developers can write their own providers by inheriting from AIProviderBase, implementing a task interface, adding an asset menu path under Meta/AI/Provider Assets and saving the provider as an asset in the project; Meta states that providers are discovered automatically at install time. As examples of what a custom provider can connect, Meta names a custom REST API or WebSocket inference service and new model types such as segmentation, diffusion or emotion recognition. A provider can also use its own network transport instead of the built-in HttpTransport, for example for gRPC or WebSocket backends.[19] Fine-tuned models can be used by exporting them to ONNX or serving them over HTTP behind a custom provider.[20]
Blocks
Meta's Building Blocks overview, updated on 8 September 2026, lists six blocks in its AI category, two of which are also listed under Passthrough, and gives v83 as the version in which all six were introduced. It places the older Dictation and Speak Text blocks, introduced in v65, in a separate Voice category.[8] The components of the four core blocks are described on Meta's Agents and Building Blocks page.[4]
| Block | Category | Purpose (Meta's description, summarized) | Main components |
|---|---|---|---|
| Passthrough Camera Access | Passthrough, AI | Direct access to the headset's camera feed for computer vision and image analysis | See Passthrough Camera API |
| Passthrough Camera Visualizer | Passthrough, AI | Shows the raw camera feed on screen for debugging | See Passthrough Camera API |
| Object Detection | AI | Returns bounding boxes for real-world objects in the camera feed | ObjectDetectionAgent, ObjectDetectionVisualizer |
| Large Language Models | AI | Multimodal models that process text, images and video to hold conversations, answer questions or analyze visual content | LlmAgent, LlmAgentHelper |
| Speech to Text | AI | Converts the user's speech into text that an LLM can interpret | SpeechToTextAgent |
| Text to Speech | AI | Generates voice audio clips from text so that an app can speak its responses | TextToSpeechAgent |
The image segmentation block from version 85.0 has no row of its own in the overview table. Meta's on-device documentation lists image segmentation as a provider mode and names an ImageSegmentationAgent.[13][21]
Object detection
The ObjectDetectionAgent takes a frame from the passthrough camera, sends the texture to its provider and receives bounding boxes, class labels and confidence scores, which the ObjectDetectionVisualizer draws in world space.[4] According to Meta's FAQ, the agent accepts both a WebCamTexture and the PassthroughCameraAccess component as input.[20] When the block is added with its visualizer, a DepthTextureAccess class is imported as well, so that 2D detections can be turned into 3D positions. DepthTextureAccess copies the per-eye environment depth textures that the EnvironmentDepthManager component retrieves from the headset into a CPU-readable array, and the agent maps each box to world space with the latest depth data. Meta calls casting a ray from the camera into the scene a still valid alternative, but says the blocks do not use it because doing so every frame and for each point of a box can be imprecise and expensive, and a head movement between inference and output can make the ray hit the wrong point.[4] Since version 85, a field called detectEveryNFrames (0 to 120) sets how often detection runs: 0 runs it only when triggered manually, 1 runs it every frame and higher values skip frames.[4]
Detection models often output several overlapping boxes for one object; non-maximum suppression (NMS) filters them so that only the most confident boxes remain. Meta supplies a GPU implementation (GpuNMS.cs and NMSCompute.compute) because the CPU-based NMS operator in some ONNX models, such as YoloX, can cause performance bottlenecks and frame drops, and on the GPUCompute backend the detections would otherwise not be filtered at all. For YOLO-style models with raw output tensors, the provider asset can convert the outputs into the tensors the blocks expect and optionally quantize the model.[21]
Language models
The LlmAgent manages text conversations and multimodal conversations that combine images and text. Its SendPromptAsync() method forwards the user's input to the provider, which streams text tokens back, and the agent raises onResponseReceived and onStreamingDelta events that other agents, such as the text-to-speech agent, can subscribe to. LlmAgentHelper connects the LLM agent to the speech agents or to other custom logic.[4] Meta lists Llama 4 Maverick, OpenAI's GPT-4o and compatible vision-language models hosted on Hugging Face as models that accept images; when a provider implements the chat interface, textures are encoded automatically and sent alongside the text prompt. The agent raises onImageCaptured when it captures a passthrough or debug image for a multimodal prompt.[4]
Speech to text and text to speech
The SpeechToTextAgent records microphone input through Unity's Microphone API or receives an audio clip, sends the encoded audio to a provider such as OpenAI or ElevenLabs, and raises onTranscript with the recognized text. Meta lists real-time and clip-based transcription, streaming recognition with some providers and an optional language code for multilingual setups.[4][20] The TextToSpeechAgent receives text through SpeakText(), requests audio from a provider such as ElevenLabs or OpenAI, and either plays the returned or streamed AudioClip or exposes it for custom playback. It raises onClipReady, onSpeakStarting and onSpeakFinished events.[4]
Combining agents
All agents share an event-driven interface and can be connected in the Unity Inspector or in code. Meta's documentation gives two example chains: microphone input passing through SpeechToTextAgent, LlmAgent and TextToSpeechAgent to audio output, and a camera frame passing through ObjectDetectionAgent into LlmAgent, which receives the detections as context, and on to TextToSpeechAgent.[4] Events that depend on the Mixed Reality Utility Kit (MRUK), such as OnBoxesUpdated, are wrapped in compile guards and exist only when that package is installed and the passthrough camera is available.[4]
Providers
The provider assets listed in Meta's documentation as of its 8 July 2026 update are:[7]
| Provider asset | Inference type | Models or capabilities listed by Meta |
|---|---|---|
| LlamaApiProvider | Cloud | Meta's Llama family, chat and multimodal variants |
| OpenAIProvider | Cloud | gpt-5 and gpt-4o (chat and vision), whisper-1 (speech-to-text), tts-1 and tts-1-hd (text-to-speech) |
| HuggingFaceProvider | Cloud | Any model hosted on Hugging Face, for example facebook/detr-resnet-101 or Llama models |
| ElevenLabsProvider | Cloud | Text-to-speech and speech-to-text (Scribe); reads voices and models from the developer's ElevenLabs account |
| RoboflowProvider | Cloud or local server | Roboflow object detection and image segmentation models |
| OllamaProvider | Local | Models installed in an Ollama daemon (at localhost:11434 in Meta's setup instructions), for example llama3, llava:latest or gemma3 |
| UnityInferenceEngineProvider | On-device | Object detection, image segmentation and chat with small language models through the Unity Inference Engine |
Meta's overview recommends the Unity Inference Engine or Hugging Face for real-time object detection; the Llama API, OpenAI, Ollama or Hugging Face for language and vision requests; and OpenAI or ElevenLabs for speech.[3] The documentation warns that models from cloud providers "may not always be available on the provider's servers" and advises checking availability before using them.[7] The page's version of 11 December 2025 listed a ReplicateProvider for community-hosted models and no Roboflow provider.[22]
On-device inference
Meta's documentation describes the on-device runtime as the "Unity Inference Engine (formerly Unity Sentis)", which runs ONNX or Sentis-format models on CPU or GPU backends.[21] Unity itself renamed the package back to Sentis in version 2.4.0 of 22 October 2025.[23] As reasons to run models on the headset, Meta lists offline operation, lower latency, privacy (passthrough images do not leave the device) and performance that does not depend on network or server load.[21] The UnityInferenceEngineProvider has three modes: object detection, image segmentation and chat with small language models. Its settings cover the backend (CPU or GPUCompute), Split Over Frames with a number of model layers to run per frame, an optional GPU NMS compute shader and a file of class labels.[21]
Meta provides an editor window (under Meta, Tools, Unity Inference Engine) that imports an ONNX model, cleans it up, optionally quantizes its weights and exports an optimized .sentis asset. Meta describes Float16 quantization as half precision that "preserves most accuracy" and Uint8 as "highly compact" with a possible slight loss of accuracy. Large models can also be serialized to the StreamingAssets folder to speed up loading.[21] The first run of a model allocates buffers, compiles GPU kernels and uploads weights, which Meta says "can cause a one-time delay of several seconds at startup", so the documentation recommends a warm-up inference during loading or splash screens. The LLM, object detection and image segmentation agents warm up assigned on-device providers when they initialize.[21]
For on-device chat, Meta lists SmolLM (135 million and 360 million parameters), Qwen 0.5B, Phi-2 and Phi-3 as supported decoder-only transformer models. A model is converted to .sentis and configured with its tokenizer vocabulary, merges and configuration files, a chat template and architecture fields. Meta's table gives a speed of 1 to 5 seconds for the blocking execution mode, which "may freeze frames", and 5 to 15 seconds for the non-blocking mode, whose effect on the frame rate it lists as smooth. The default limits are 100 generated tokens and a 512-token prompt.[21]
Requirements
Meta's system recommendations for the AI Building Blocks are Unity 6 or newer, a Meta Quest 3 or Quest 3S, version 83 or later of both the Meta XR Core SDK and the Meta XR MR Utility Kit, which Meta lists "for Passthrough Camera support", and a stable internet connection when inference runs in the cloud or on a local machine over HTTP.[3] Since Core SDK 203.0, the SDK as a whole requires Unity 6000.0.66f2 or later.[17] Version 83.0 of the MR Utility Kit, released the same day as Core SDK 83.0, also changed Passthrough Camera Access, for example adding Vulkan support and frame rates of up to 60 frames per second.[24] Cloud providers need an API key for the respective service; Meta's FAQ notes that some models, such as gpt-4o, require a premium subscription or API credits.[7][20]
Use in VR development
At Connect 2025, Meta wrote that object detection "can form the basis of reacting in real time to objects or changes in the environment" and that the LLM block can be used to "interpret and interact with your environment".[6] At its Unite 2025 booth, guided demos showed how to drag and drop features including passthrough, hand tracking and AI into a Unity project.[9] At GDC 2026, Meta presented the blocks for adding "scene-aware" behavior that combines camera access with object detection and LLM prompts, and for adding conversation and voice control.[14][15]
Limitations
Meta's documentation states several limitations. Cloud models may not always be available on providers' servers, and cloud inference and local inference over HTTP need a stable network connection.[3][7] To reduce cost, the FAQ suggests using local models (Ollama or the Unity Inference Engine) during development before switching to cloud providers for production builds, and avoiding inference on every frame unless necessary; it also warns against shipping API credentials in a production build or a GitHub repository.[20] On-device inference carries what Meta calls a "significant performance impact", and Meta's figures for on-device chat range from 1 to 15 seconds depending on the execution mode.[7][21]
The documentation of July 2026 also describes two open issues. On the GPUCompute backend, object detection can cause performance spikes because it can require synchronization between CPU and GPU while results are filtered and converted into 3D bounding boxes, a bottleneck that Meta says will be resolved in a future release.[21] Bounding boxes can also be offset with some passthrough camera resolutions and aspect ratios. Meta says it has a fix ready for the upcoming release and advises developers who see an offset to set the PassthroughCameraAccess resolution to 1280 by 960, the earlier default.[4]
Current status
As of November 2026 the AI Building Blocks are part of the Meta XR Core SDK. The newest entry in Meta's Core SDK release notes is version 207.0, dated 22 September 2026.[25] The Unity Asset Store lists a later version, 207.0.1, released on 12 October 2026, whose Core SDK note concerns other tools: "Added Muse Code setup support to AI tools, including XR Operator and AI Agent Bridge".[1] The release notes for versions 203.0, 205.0 (22 July 2026) and 207.0 list no changes to the AI Building Blocks. These releases added separate AI-related tools, such as the AI Runtime Optimizer Tool in 203.0 and Meta XR Operator in 205.0, which lets AI coding agents observe, interact with and verify a running app.[17][26][25] The AI Building Blocks documentation was last updated on 8 July 2026. Meta's Building Blocks overview of 8 September 2026 lists the six AI-category blocks in its main table; the only block it marks as experimental is the separately listed Gaze Interaction block.[3][8]
See also
References
- ↑ 1.0 1.1 "Meta XR SDK". Unity Asset Store. Unity Technologies. https://assetstore.unity.com/packages/sdk/meta-xr-sdk-9022845. Retrieved 2026-11-01.
- ↑ "Meta Platform Technologies SDK License Agreement". Meta Horizon OS Developers. Meta. 2022-10-25. https://developers.meta.com/horizon/licenses/oculussdk/. Retrieved 2026-11-01.
- ↑ 3.00 3.01 3.02 3.03 3.04 3.05 3.06 3.07 3.08 3.09 3.10 3.11 "AI Building Blocks - Overview". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-building-blocks-overview/. Retrieved 2026-11-01.
- ↑ 4.00 4.01 4.02 4.03 4.04 4.05 4.06 4.07 4.08 4.09 4.10 4.11 4.12 4.13 4.14 4.15 4.16 4.17 "Agents and Building Blocks". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-agents/. Retrieved 2026-11-01.
- ↑ 5.0 5.1 5.2 "Meta XR Core SDK (UPM), version 83.0 release notes". Meta Horizon OS Developers. Meta. 2025-12-10. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/83.0/. Retrieved 2026-11-01.
- ↑ 6.0 6.1 6.2 6.3 "Meta Connect 2025: Fueling the Future of VR". Meta Horizon OS Developers Blog. Meta. 2025-09-18. https://developers.meta.com/vr/blog/meta-connect-2025-fueling-the-future-of-vr/. Retrieved 2026-11-01.
- ↑ 7.00 7.01 7.02 7.03 7.04 7.05 7.06 7.07 7.08 7.09 7.10 7.11 7.12 7.13 "Providers and Inference Types". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-providers/. Retrieved 2026-11-01.
- ↑ 8.0 8.1 8.2 "Building Blocks Overview". Meta Horizon OS Developers. Meta. 2026-09-08. https://developers.meta.com/vr/documentation/unity/unity-building-blocks-overview/. Retrieved 2026-11-01.
- ↑ 9.0 9.1 "Unity, Meta Horizon OS, and the Future of VR at Unite 2025". Meta Horizon OS Developers Blog. Meta. 2025-11-25. https://developers.meta.com/vr/blog/unity-meta-horizon-os-the-future-of-vr-unite-2025/. Retrieved 2026-11-01.
- ↑ "AI Building Blocks - Overview". Meta Horizon OS Developers. Meta. 2025-12-11. Archived from the original on 2026-01-20. https://web.archive.org/web/20260120153952/https://developers.meta.com/horizon/documentation/unity/unity-ai-building-blocks-overview/. Retrieved 2026-11-01.
- ↑ Dalton Pakkala, Sam Sklar (2025-12-11). "ElevenLabs is partnering with Meta to power expressive, scalable audio". ElevenLabs Blog. ElevenLabs. https://elevenlabs.io/blog/meta. Retrieved 2026-11-01.
- ↑ Maria Stasimioti (2025-12-17). "ElevenLabs Wins Meta as Key Account". Slator. Archived from the original on 2026-09-18. https://web.archive.org/web/20260918073453/https://slator.com/elevenlabs-wins-meta-as-key-account/. Retrieved 2026-11-01.
- ↑ 13.0 13.1 13.2 "Meta XR Core SDK (UPM), version 85.0 release notes". Meta Horizon OS Developers. Meta. 2026-02-10. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/85.0/. Retrieved 2026-11-01.
- ↑ 14.0 14.1 "Faster Builds, Smarter Discovery, and the LiveOps Playbook: What to Know After GDC Day 2". Meta Horizon OS Developers Blog. Meta. 2026-03-11. https://developers.meta.com/vr/blog/gdc-2026-day-2-faster-builds-discovery-what-to-know/. Retrieved 2026-11-01.
- ↑ 15.0 15.1 "GDC 2026 Highlights: What's Next on Meta Horizon OS". Meta Horizon OS Developers Blog. Meta. 2026-03-13. https://developers.meta.com/vr/blog/gdc-2026-highlights-whats-next-meta-horizon-os/. Retrieved 2026-11-01.
- ↑ 16.0 16.1 "Meta XR Core SDK (UPM), version 201.0 release notes". Meta Horizon OS Developers. Meta. 2026-04-15. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/201.0/. Retrieved 2026-11-01.
- ↑ 17.0 17.1 17.2 "Meta XR Core SDK (UPM), version 203.0 release notes". Meta Horizon OS Developers. Meta. 2026-06-03. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/203.0/. Retrieved 2026-11-01.
- ↑ "What Unity Developers Learned About Building for Meta Quest at Unite Seoul 2026". Meta Horizon OS Developers Blog. Meta. 2026-07-22. https://developers.meta.com/vr/blog/unite-seoul-2026-meta-quest/. Retrieved 2026-11-01.
- ↑ 19.0 19.1 19.2 "Adding New Providers". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-add-new-provider/. Retrieved 2026-11-01.
- ↑ 20.0 20.1 20.2 20.3 20.4 20.5 "Troubleshooting and FAQ". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-faq/. Retrieved 2026-11-01.
- ↑ 21.00 21.01 21.02 21.03 21.04 21.05 21.06 21.07 21.08 21.09 "Unity Inference Engine". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/vr/documentation/unity/unity-ai-unity-inference-engine/. Retrieved 2026-11-01.
- ↑ "Providers and Inference Types". Meta Horizon OS Developers. Meta. 2025-12-11. Archived from the original on 2026-01-20. https://web.archive.org/web/20260120141044/https://developers.meta.com/horizon/documentation/unity/unity-ai-providers/. Retrieved 2026-11-01.
- ↑ "Changelog - Sentis 2.6". Unity Documentation. https://docs.unity3d.com/Packages/[email protected]/changelog/CHANGELOG.html. Retrieved 2026-11-01.
- ↑ "Meta XR MR Utility Kit (UPM), version 83.0 release notes". Meta Horizon OS Developers. Meta. 2025-12-10. https://developers.meta.com/vr/downloads/package/meta-xr-mr-utility-kit-upm/83.0/. Retrieved 2026-11-01.
- ↑ 25.0 25.1 "Meta XR Core SDK (UPM), version 207.0 release notes". Meta Horizon OS Developers. Meta. 2026-09-22. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/207.0/. Retrieved 2026-11-01.
- ↑ "Meta XR Core SDK (UPM), version 205.0 release notes". Meta Horizon OS Developers. Meta. 2026-07-22. https://developers.meta.com/vr/downloads/package/meta-xr-core-sdk/205.0/. Retrieved 2026-11-01.