Meta models
Meta Platforms builds the Llama open-weights models and, through Meta Superintelligence Labs, the Muse model family, alongside FAIR research models such as Segment Anything.
Menlo Park, United StatesFounded 2004meta.com
- Releases
- 44
- Open weights
- 35
- Latest
- Nov 09, 2026
November 2026
2 releasesMuse Ember
Muse Ember is the smallest and most efficient model in Meta's Muse family, designed for complex reasoning execution, with native understanding of text, images, audio, and video.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms
- Available at launch on the Meta Model API and in Muse Code
- 1M-token context window and up to 512K output tokens
- The first model trained using Muse Wonder Full
- Offered on Google Cloud and Oracle Cloud alongside Muse Spark and Muse Wonder
Specialized models for Muse and Meta AI
Meta introduced specialized versions of its existing models, trained using Muse Wonder, and began rolling them out to people using Muse and Meta AI. The announcement gave the models no separate names.
- At launch
- Consumer app
- Trained using the Muse Wonder series, which Meta described as its most capable models to date
- Brought Muse Wonder's Active Memory and Active Focus to Muse and Meta AI
- Meta said the models were developed to multitask, adapt to mid-task clarifications, and ask sharper follow-up questions
October 2026
2 releasesMuse Spark Open
Muse Spark Open is the open-weights member of Meta Superintelligence Labs' Muse family, a natively multimodal mixture-of-experts model built for coding and agentic work with a 1M-token context window.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- Parameters
- MoE: 816B total, 42B active
- License
- Muse Model License
- At launch
- Open weights, Cloud platforms
- 816B total parameters with 42B active per token
- Model card lists text, image, video and audio input with text output
- Weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry
- Hosted by providers including Baseten, Fireworks AI, Groq and Together AI
Muse Wonder and Muse Wonder Full
Muse Wonder and Muse Wonder Full open the Muse Wonder series, Meta's most capable models, aimed at personal and professional agentic tasks with Active Memory and Active Focus features.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- At launch
- API, Consumer app
- Muse Wonder is available through the Meta Model API and Muse Code
- Muse Wonder Full is available to Muse Code subscribers only
- Meta Model API docs list a 1,048,576-token context window for Muse Wonder
- Muse Wonder is slated to power the Muse personal agent, first for subscribers
September 2026
2 releasesMuse Spark 1.3
Muse Spark 1.3 is an update to Meta's Muse Spark reasoning model that improves long-horizon agentic and coding work, follows long instructions more reliably, and adds a max reasoning setting.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- At launch
- API
- Available with max reasoning on Muse Code and the Meta Model API
- Takes fewer unnecessary turns and is less verbose in coding than Muse Spark 1.2
- Meta reports stronger adversarial robustness, including against prompt injection
- API docs list a 1,048,576-token context window; audio input is not fully supported
Muse Voice Transcribe
Muse Voice Transcribe is Meta Superintelligence Labs' first real-time audio perception model, combining streaming speech recognition, speaker diarization and endpointing in one model.
- Modalities
- audio → text
- At launch
- API, Consumer app
- Diarization for more than 20 speakers
- Trained on 70+ languages, with 25 extensively verified, and handles code-switching
- Available in the Meta Model API, Meta AI for Mac and Muse Code
- API model ID muse-voice-transcribe-1.0
August 2026
2 releasesMuse Glimmer
Muse Glimmer is a 30B open-weights model from Meta Superintelligence Labs, distilled from Muse Spark and built for always-on local agent workflows on consumer hardware.
- Modalities
- text, image → text
- Context
- 128K context
- Parameters
- 30B (dense)
- License
- Apache-2.0
- At launch
- Open weights
- Released under Apache 2.0, the most permissive license Meta has used for an open model
- Accepts interleaved text and images through a dedicated perception encoder
- Trained on Muse Spark outputs using logit distillation
- Weights on Hugging Face, including GGUF and ExecuTorch builds
Muse Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, launched alongside Muse Code, Meta's terminal coding agent, with gains in code generation, debugging and codebase understanding.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- At launch
- API
- Launched with Muse Code, a beta terminal coding agent
- Available in Muse Code and the Meta Model API with expanded global access
- Meta scaled up training compute on coding tasks
- API docs list a 1,048,576-token context window
July 2026
2 releasesMuse Spark 1.1
Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, upgrading Muse Spark in tool and computer use, coding, multi-agent orchestration and context management.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- At launch
- API, Consumer app
- Available to developers through the Meta Model API public preview
- Manages a context window of 1 million tokens
- Available in Thinking mode in the Meta AI app and on meta.ai
Muse Image
Muse Image is Meta Superintelligence Labs' first image generation model, an agentic generator and editor that uses search and coding tools, refines its own outputs and composes from multiple references.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Available in the Meta AI app, on meta.ai, and in Instagram Stories in the US at launch
- Supports multi-turn precise editing and multi-reference composition
- Outputs carry Content Seal, an invisible watermark
- Announced with Muse Video, which was only previewed
April 2026
1 releaseMuse Spark
Muse Spark is the first model from Meta Superintelligence Labs, a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration that powers Meta AI.
- Modalities
- text, image → text
- At launch
- Consumer app
- Available at launch on meta.ai and in the Meta AI app
- Contemplating mode runs multiple agents that reason in parallel
- API access opened only as a private preview for select users
- Weights were not released; Meta said it hopes to open-source future versions
March 2026
1 releaseSAM 3.1
SAM 3.1 updates Segment Anything Model 3 with Object Multiplex, a shared-memory approach to joint multi-object tracking that makes real-time video detection and tracking significantly faster.
- Modalities
- text, image, video → image, video
- License
- SAM License
- At launch
- Open weights, API
- Drop-in replacement checkpoint for SAM 3, published on Hugging Face
- Faster multi-object video tracking without sacrificing accuracy
- Offered as sam-3.1 in the Meta Model API
December 2025
1 releaseSAM Audio
SAM Audio extends the Segment Anything approach to sound, a unified model that isolates specific sounds from audio mixtures using text, visual or time-span prompts.
- Modalities
- audio, text, video → audio
- Parameters
- 500M to 3B
- License
- SAM License
- At launch
- Open weights
- Prompts can be text, clicks on sounding objects in video, or marked time spans
- Released alongside the PE-AV perception encoder
- Audio prompts are not supported
November 2025
3 releasesSAM 3
SAM 3 is a unified model for detecting, segmenting and tracking objects in images and video, adding promptable concept segmentation from short text phrases or image exemplars.
- Modalities
- text, image, video → image, video
- Parameters
- 848M
- License
- SAM License
- At launch
- Open weights
- Finds all instances of a concept from a noun-phrase or exemplar prompt
- Weights, the SA-Co benchmark and a paper were released
- Available to try in the Segment Anything Playground
SAM 3D (Objects and Body)
SAM 3D is a pair of models that reconstruct 3D objects, scenes and human bodies from a single image: SAM 3D Objects for objects and scenes and SAM 3D Body for human pose and shape.
- Modalities
- image → 3d
- License
- SAM License
- At launch
- Open weights
- Checkpoints and inference code released on GitHub
- MHR, the parametric human model used by SAM 3D Body, is released under a permissive commercial license
- Available to try in the Segment Anything Playground
Omnilingual ASR
Omnilingual ASR is a suite of open speech recognition models from Meta FAIR covering more than 1,600 languages, including 500 low-resource languages not previously transcribed by AI.
- Modalities
- audio → text
- Parameters
- 300M to 7B
- License
- Apache-2.0
- At launch
- Open weights
- Models range from 300M versions for low-power devices to 7B models
- Can extend to new languages from a few in-context audio-text examples
- Released with the Omnilingual ASR Corpus covering 350 underserved languages
September 2025
1 releaseCode World Model (CWM)
Code World Model is a 32B dense, decoder-only open-weights LLM from Meta FAIR, released to support research on code generation with world models.
- Modalities
- text → text
- Context
- 128K context
- Parameters
- 32B
- License
- FAIR Non-Commercial Research License
- At launch
- Open weights
- Dense decoder-only model trained with up to 131k tokens of context
- Checkpoints released after mid-training, SFT and RL
- Intended for non-commercial research, not as a chat assistant
April 2025
4 releasesLlama Guard 4
Llama Guard 4 is a 12B natively multimodal safety classifier that moderates text and multi-image inputs and model responses, acting as a unified safeguard across modalities.
- Modalities
- text, image → text
- Parameters
- 12B
- License
- Llama 4 Community License
- At launch
- Open weights, API
- Dense model pruned from the Llama 4 Scout pretrained model
- Classifies content against 14 MLCommons-based hazard categories
- Also offered on the Llama API, then in limited preview
Llama Prompt Guard 2
Llama Prompt Guard 2 updates Meta's prompt attack classifier with an improved 86M model and a smaller 22M model for detecting prompt injection and jailbreak attempts.
- Modalities
- text → text
- Context
- 1K context
- Parameters
- 86M, 22M
- License
- Llama 4 Community License
- At launch
- Open weights
- The 22M model cuts latency and compute costs by up to 75% versus the 86M model
- 86M is based on mDeBERTa-base; 22M on DeBERTa-xsmall
- Both support a 512-token context window
Llama 4 Maverick
Llama 4 Maverick is a natively multimodal mixture-of-experts model with 17B active parameters and 128 experts, released with Llama 4 Scout as the first Llama 4 models.
- Modalities
- text, image → text
- Context
- 1M context
- Parameters
- MoE: 400B total, 17B active, 128 experts
- License
- Llama 4 Community License
- At launch
- Open weights, Cloud platforms, Consumer app
- 400B total parameters; runs on a single NVIDIA H100 DGX host
- Model card lists a 1M-token context window and 12 supported languages
- Knowledge cutoff of August 2024
- Powers Meta AI in WhatsApp, Messenger, Instagram Direct and meta.ai
Llama 4 Scout
Llama 4 Scout is a natively multimodal mixture-of-experts model with 17B active parameters and 16 experts, notable for a 10 million token context window.
- Modalities
- text, image → text
- Context
- 10M context
- Parameters
- MoE: 109B total, 17B active, 16 experts
- License
- Llama 4 Community License
- At launch
- Open weights, Cloud platforms
- 109B total parameters; fits on a single NVIDIA H100 GPU with Int4 quantization
- Accepts multilingual text and image input
- Llama 4 Behemoth was previewed but not released because it was still training
December 2024
1 releaseLlama 3.3 70B
Llama 3.3 is a 70B instruction-tuned multilingual text model with a 128K context window, released as a single model in the Llama 3 line.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 70B
- License
- Llama 3.3 Community License
- At launch
- Open weights, Cloud platforms
- Officially supports eight languages, including English, German, Hindi and Thai
- Pretrained on about 15 trillion tokens with a December 2023 cutoff
- Text-only input; outputs text and code
October 2024
1 releaseLlama 3.2 1B and 3B quantized
Meta released quantized versions of Llama 3.2 1B and 3B, its first lightweight quantized Llama models, built with QAT plus LoRA and with SpinQuant to run on mobile devices.
- Modalities
- text → text
- Parameters
- 1B, 3B
- License
- Llama 3.2 Community License
- At launch
- Open weights
- 2 to 4 times faster than the original BF16 models on average
- Average 56% smaller model size and 41% lower memory use
- Model card lists an 8K context length for the quantized models
September 2024
5 releasesSAM 2.1
SAM 2.1 is an updated set of Segment Anything Model 2 checkpoints with stronger performance, released with the SAM 2 training code and web demo code as the SAM 2.1 Developer Suite.
- Modalities
- image, video → image, video
- License
- Apache-2.0
- At launch
- Open weights
- New checkpoints require the latest SAM 2 repository code
- Training and fine-tuning code released
- Front-end and back-end code for the SAM 2 web demo released
Llama 3.2 11B and 90B Vision
Llama 3.2 Vision adds image understanding to Llama with 11B and 90B parameter models that take text and images as input and produce text output.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 11B, 90B
- License
- Llama 3.2 Community License
- At launch
- Open weights, Cloud platforms
- Model card lists a 128K context length
- English is the only supported language for image plus text tasks
- Released with the first official Llama Stack distributions
Llama 3.2 1B and 3B
Llama 3.2 1B and 3B are lightweight text-only models designed for edge and mobile devices, supporting a 128K token context window.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 1B, 3B
- License
- Llama 3.2 Community License
- At launch
- Open weights, Cloud platforms
- Officially supports eight languages
- Pretraining data cutoff of December 2023
- Released alongside the Llama 3.2 Vision models
Llama Guard 3 11B Vision
Llama Guard 3 11B Vision is a safety classifier fine-tuned from Llama 3.2 11B that filters text plus image prompts and text responses, supporting Llama 3.2's image understanding.
- Modalities
- text, image → text
- Parameters
- 11B
- License
- Llama 3.2 Community License
- At launch
- Open weights
- Fine-tuned from the Llama 3.2 11B pretrained model
- Predicts safety labels across 13 MLCommons-based hazard categories
Llama Guard 3 1B
Llama Guard 3 1B is a compact content safety classifier based on Llama 3.2 1B, with a pruned and quantized version for deployment on mobile devices.
- Modalities
- text → text
- Parameters
- 1B
- License
- Llama 3.2 Community License
- At launch
- Open weights
- Pruned and quantized version shrinks from 2,858 MB to 438 MB
- Covers 13 hazard categories in eight languages
July 2024
4 releasesSAM 2
SAM 2 is Meta's next-generation Segment Anything Model, a unified model for real-time promptable object segmentation in both images and videos.
- Modalities
- image, video → image, video
- License
- Apache-2.0
- At launch
- Open weights
- Runs at about 44 frames per second
- Released with the SA-V dataset of about 51,000 videos under CC BY 4.0
- Treats an image as a single-frame video so one architecture handles both
Llama 3.1
Llama 3.1 introduced the 405B model, which Meta described as the largest openly available foundation model, plus upgraded multilingual 8B and 70B models, all with a 128K context window.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 8B, 70B, 405B
- License
- Llama 3.1 Community License
- At launch
- Open weights, Cloud platforms
- Supports eight languages
- License changed to allow using Llama outputs to improve other models
- 405B trained on over 15 trillion tokens
- Over 25 partners, including AWS, Azure and Google Cloud, offered services on day one
Llama Guard 3
Llama Guard 3 is a multilingual content safety classifier fine-tuned from Llama 3.1 8B, released with Llama 3.1 and optimized for search and code interpreter tool calls.
- Modalities
- text → text
- Parameters
- 8B
- License
- Llama 3.1 Community License
- At launch
- Open weights
- Predicts safety labels across 14 categories, adding Code Interpreter Abuse
- Supports eight languages
Prompt Guard
Prompt Guard is a small classifier released with Llama 3.1 that labels inputs as benign, prompt injection or jailbreak, built to filter prompt attacks in LLM applications.
- Modalities
- text → text
- Context
- 1K context
- Parameters
- 86M
- License
- Llama 3.1 Community License
- At launch
- Open weights
- Fine-tuned from mDeBERTa-v3-base with 86M backbone parameters
- 512-token context window; longer inputs are split into segments
June 2024
1 releaseChameleon
Chameleon is a mixed-modal early-fusion model family from Meta FAIR; Meta released key components of the 7B and 34B models for research with text-only output.
- Modalities
- text, image → text
- Parameters
- 7B, 34B
- License
- Chameleon Research License
- At launch
- Open weights
- Accepts mixed text and image input
- Released under a research-only license
- The Chameleon image generation model was not released
April 2024
2 releasesLlama 3
Llama 3 is Meta's third-generation open model family, released as 8B and 70B pretrained and instruction-tuned models trained on over 15 trillion tokens.
- Modalities
- text → text
- Context
- 8K context
- Parameters
- 8B, 70B
- License
- Meta Llama 3 Community License
- At launch
- Open weights, Cloud platforms
- Trained on sequences of 8,192 tokens
- Released with Llama Guard 2, Code Shield and CyberSec Eval 2
- Meta said models over 400B parameters were still in training
Llama Guard 2
Llama Guard 2 is an 8B Llama 3 based input and output safeguard model released with Llama 3, using the MLCommons hazard taxonomy.
- Modalities
- text → text
- Parameters
- 8B
- License
- Meta Llama 3 Community License
- At launch
- Open weights
- Predicts safety labels on 11 of the 13 MLCommons hazard categories
- Can be fine-tuned for different applications
January 2024
1 releaseCode Llama 70B
Code Llama 70B is the largest model in the Code Llama family, released in base, Python and Instruct versions and free for research and commercial use.
- Modalities
- text → text
- Parameters
- 70B
- License
- Llama 2 Community License
- At launch
- Open weights
- Same three versions as the original release: Code Llama, Code Llama - Python and Code Llama - Instruct
- Free for research and commercial use
December 2023
1 releaseLlama Guard
Llama Guard is a 7B Llama 2 based safety classifier for filtering LLM inputs and outputs, released as part of Meta's Purple Llama trust and safety project.
- Modalities
- text → text
- Parameters
- 7B
- License
- Llama 2 Community License
- At launch
- Open weights
- Instruction-tuned from Llama 2 7B
- Covers six harm categories in its original taxonomy
- Launched with CyberSec Eval cybersecurity benchmarks under Purple Llama
November 2023
1 releaseSeamless (SeamlessM4T v2, SeamlessExpressive, SeamlessStreaming)
Seamless is a suite of four translation models built on SeamlessM4T v2, adding expression-preserving speech translation and streaming translation with about two seconds of latency.
- Modalities
- text, audio → text, audio
- Parameters
- 2.3B (SeamlessM4T v2 Large), 2.5B (SeamlessStreaming)
- License
- CC-BY-NC-4.0 (SeamlessM4T v2, SeamlessStreaming); Seamless License (SeamlessExpressive)
- At launch
- Open weights
- SeamlessExpressive preserves expression across English, Spanish, German, French, Italian and Chinese
- SeamlessStreaming supports speech recognition and speech-to-text translation for nearly 100 languages
- All four models released publicly for researchers
August 2023
3 releasesCode Llama
Code Llama is a family of code-specialized Llama 2 models, released in base, Python and Instruct variants for code generation and completion.
- Modalities
- text → text
- Context
- 100K context
- Parameters
- 7B, 13B, 34B
- License
- Llama 2 Community License
- At launch
- Open weights
- Stable generations with up to 100,000 tokens of context
- Code Llama - Python was further fine-tuned on 100B tokens of Python code
- Released under the same community license as Llama 2
SeamlessM4T
SeamlessM4T is a single multimodal model for speech and text translation and speech recognition, covering nearly 100 input languages and 35 speech output languages plus English.
- Modalities
- text, audio → text, audio
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Speech recognition for nearly 100 languages
- Speech-to-speech translation into 35 languages plus English
- Released publicly under CC BY-NC 4.0
AudioCraft (MusicGen and AudioGen)
Meta open-sourced AudioCraft, its generative audio framework, releasing model weights and code for MusicGen text-to-music, AudioGen text-to-sound and an improved EnCodec decoder.
- Modalities
- text → audio
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Model weights released under CC-BY-NC 4.0 for research use
- Training code released under the MIT license
July 2023
1 releaseLlama 2
Llama 2 is Meta's second-generation open model family, released free for research and commercial use with pretrained and chat fine-tuned versions in 7B, 13B and 70B sizes.
- Modalities
- text → text
- Context
- 4K context
- Parameters
- 7B, 13B, 70B
- License
- Llama 2 Community License
- At launch
- Open weights, Cloud platforms
- Pretrained on 2 trillion tokens of publicly available data
- Launched with Microsoft as preferred partner, in the Azure AI model catalog
- Also available through AWS and Hugging Face
April 2023
1 releaseSegment Anything Model (SAM)
Segment Anything Model is a promptable image segmentation foundation model that can segment objects from clicks or boxes and transfer zero-shot to new image domains.
- Modalities
- image → image
- License
- Apache-2.0
- At launch
- Open weights
- Released with SA-1B, over 1.1 billion masks on about 11 million images
- Produces a segment in about 50 milliseconds in a browser once the image embedding is computed
February 2023
1 releaseLLaMA
LLaMA is Meta's first family of foundation language models, released to researchers under a noncommercial license in four sizes from 7B to 65B parameters.
- Modalities
- text → text
- Parameters
- 7B, 13B, 33B, 65B
- License
- Non-commercial bespoke license
- At launch
- Open weights
- LLaMA 65B and 33B trained on 1.4 trillion tokens; 7B on one trillion
- Access granted case by case to academic researchers and selected organizations
Cite and reuse this data
These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.
Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".
Citation
The AI Engineering Editors. "Meta models: release history and timeline." The AI Engineering, updated Nov 09, 2026. https://theaiengineering.com/releases/meta
HTML link
<a href="https://theaiengineering.com/releases/meta">Meta models: release history and timeline</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering
Other labs
Google106
Alibaba Qwen84
OpenAI79Mistral AI59
- Z.ai43
Microsoft42Tencent42
- ByteDance Seed40
- DeepSeek38
- Liquid AI34
Cohere31- IBM30
xAI30- NVIDIA29
MiniMax28
Stability AI28- Anthropic26
Ai223StepFun22
Amazon21- Baidu21
- ElevenLabs21
Black Forest Labs18
- Kuaishou16
- Meituan16
Runway16- Moonshot AI15
Midjourney14Salesforce13
- Xiaomi13
Nous Research11- Perplexity10
AI21 Labs8- Apple8
Ideogram8
01.AI7
Hugging Face7
Lightricks7Luma AI7
Reka7- Suno7
- Databricks5
Leonardo.Ai5- Pika5
Recraft5
Inflection AI3MBZUAI IFM3
- Snowflake3
Thinking Machines Lab3
TypeSafe AI2- Cognition1
Genmo1Reflection AI1
Ricursive Intelligence1