Every Meta AI model, compared

The Muse family: the current line

Muse is where Meta's model work lives now. The name does triple duty, so a quick untangling first. Muse is the consumer personal agent app, launched September 8, 2026. Muse Code is the terminal coding agent, not a model. The models below are the actual models, and they form a ladder: small and cheap at the bottom, frontier at the top.

Muse Ember November 9, 2026

The smallest Muse model, and the newest. Meta built it for the token-heavy work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. It is multimodal, reading text, images, and audio, including screenshots, charts, scanned documents, voice memos, and meeting recordings. Price: $0.10 in, $0.50 out per million tokens, per Meta's announcement. That price had not reached Meta's own pricing page yet at the time of writing, so treat it as announcement pricing. Available on the Meta Model API and in Muse Code. Meta's own framing: Wonder plans, Ember handles the volume. In its announcement Meta also positioned Ember against Anthropic's Haiku and OpenAI's Luna on cost per task, using Artificial Analysis benchmarks.

Muse Spark April to September 2026

The workhorse line, in four versions. Spark 1.0 arrived April 8, 2026 as the first Muse model and the engine behind the Meta AI app. Spark 1.1 (July 9) opened the Meta Model API to the public in preview and brought a 1M-token context window. Spark 1.2 (August 5) was tuned for coding and launched the Muse Code beta. Spark 1.3 (September 2) is the current flagship of the line: long-horizon agentic and coding work, natively multimodal across text, image, and video, with two reasoning settings. Meta says 1.3 makes about 20% fewer tool calls and uses about 25% fewer tokens than 1.2. Price: $1.25 in, $0.15 cached, $4.25 out. There is also a contributor tier at $0.10 in and $0.20 out, where Meta trains on your data. In April 2026 Spark replaced Llama 4 as the engine behind Meta AI, which is the clearest sign of where Meta's priorities moved.

Muse Glimmer 30B August 10, 2026

The small open-weight model, about 29.6B parameters including a roughly 2B vision encoder, distilled from Spark, with a 128K context window. Apache 2.0 license, free weights on Hugging Face, no API key needed. Built for local agents on a single consumer GPU or a Mac.

Muse Wonder and Wonder Full October 12, 2026

The frontier flagships, Meta's largest and most capable models, with a 1M-token context window and two features called Active Memory and Active Focus, built for long-horizon personal and professional agent work. Wonder costs $4 in, $0.20 cached, $20 out. Wonder Full is the most capable variant and is not sold by the token at all: it is only available through Muse Code subscriptions, Power at $50 a month and Maximum at $200 a month.

Muse Spark Open October 15, 2026

The open-weight version of Spark: 816B total parameters with about 42B active per token, a mixture-of-experts design with 256 experts, 62 layers, and a 1M-token context window. Free weights under Meta's Muse Model License, on Hugging Face and the major clouds. Meta's guidance says it runs on two Mac Studios, which tells you it is open but not light.

Muse Image July 7, 2026

The media model in the family. It generates and edits images, and it reasons before it renders, calling web search and code tools in the middle of generation and refining its own output. Consumer launch July 7, API launch August 28. Price: a flat $0.01 per image. Free for everyday use in the Meta AI app.

Muse Video Previewed July 7, 2026

Text to video with synchronized audio. Still in limited preview, not released.

Muse Voice Transcribe

Speech to text with speaker attribution, $0.18 per hour of audio on the Meta Model API. The release date is not pinned down publicly, so treat the date as unconfirmed even though the price is confirmed on Meta's pricing page.

The price ladder is the cleanest way to read the family. Ember at $0.10 in, Spark at $1.25, Wonder at $4, and Wonder Full off the token meter entirely. Ember is about 12.5 times cheaper than Spark on input and 40 times cheaper than Wonder. That is the whole design: a heavy model sets the rules, a light model applies them across tens of thousands of cases.

What is not a model

Three things with Muse in the name are not models. Muse, the agent app, runs on Spark and later Wonder for subscribers. Muse Code, the terminal coding agent, runs Spark, Wonder, and Ember. And a November 6 update rolled out specialized Wonder-trained versions for Muse and Meta AI, which is a tuning update, not a new named model.

The Llama family: the predecessor, now legacy

Llama is the family that made Meta the open model company, and it is effectively paused. The newest Llama generation is Llama 4, from April 5, 2025. There is no Llama 5 and no 2026 Llama release.

LLaMA February 24, 2023

The original: 7B, 13B, 33B, and 65B, research only, non-commercial license. The weights leaked within about two weeks. Its lasting point was that small models trained on far more data could beat much larger ones: the 13B beat GPT-3 on most benchmarks.

Llama 2 July 18, 2023

7B, 13B, and 70B, under the Llama 2 Community License, the first version that allowed commercial use, with a 700 million monthly active user cap. Added RLHF chat models and a 4K context window.

Llama 3 April 18, 2024

8B and 70B, trained on more than 15 trillion tokens, with a 128K-token vocabulary and an 8K context window.

Llama 3.1 July 23, 2024

8B, 70B, and 405B. The 405B was the first open-weight model Meta called frontier level. 128K context, eight languages, tool calling.

Llama 3.2 September 25, 2024

1B and 3B text models for on-device use, plus 11B and 90B vision models, the first multimodal Llamas. It also brought the Llama Stack toolkit.

Llama 3.3 December 6, 2024

A single 70B instruct model whose post-training gains approached 405B quality at a fraction of the serving cost. Still heavily used.

Llama 4 Scout and Maverick April 5, 2025

The first mixture-of-experts Llamas, natively multimodal. Scout is 17B active out of 109B total with a claimed 10M-token context window. Maverick is 17B active out of 400B total with a 1M-token context window. Both are still available, but Meta's own assistant has moved off them.

Llama 4 Behemoth Previewed April 5, 2025, never released

About 288B active out of roughly 2 trillion total parameters, the teacher model Scout and Maverick were distilled from. It was described as still training at preview, was repeatedly delayed, and never shipped. Meta has never formally cancelled it, so the honest status is previewed, never released, never formally cancelled. Do not read more into it than that.

The wind-down

Meta's hosted Llama API public preview shut down on July 6, 2026. The weights are unaffected and still downloadable from Meta and Hugging Face, served by third parties like AWS Bedrock, Together AI, Groq, and DeepInfra. But even the fallback hosts are thinning: Groq deprecated Maverick in March 2026 and Scout in July 2026. Llama 3.x dense models remain in heavy production use, the 8B at the edge, the 70B on a single server, the 405B in datacenters. The family is not dead. It is just no longer where Meta builds.

The research lab: vision, video, audio, and translation

Most of Meta's models never became products. They are research releases, and they are worth one section, not one post each.

Vision

Segment Anything, or SAM, arrived April 5, 2023: promptable image segmentation trained on 11 million images and 1.1 billion masks, Apache 2.0, widely used for data labeling. SAM 2 followed July 29, 2024 with real-time segmentation in images and video. SAM 3 arrived November 19, 2025 and added detection, segmentation, and tracking of every instance of a concept named by a short text phrase or image example. SAM 3.1, updated March 27, 2026, is a live product on Meta's Model API at $2.50 per 1,000 images and $0.20 per 1,000 video frames. DINOv3, released August 2025, is a 7B-parameter self-supervised vision model trained on about 1.7 billion unlabeled images, with a single frozen backbone that matches or beats specialized models on roughly 60 benchmarks. I-JEPA (June 2023) and V-JEPA (February 2024) predict masked regions in representation space instead of pixels; V-JEPA 2, from June 11, 2025, is a 1.2B-parameter video world model trained on more than a million hours of video, with an action-conditioned variant that does zero-shot robot pick-and-place planning. ImageBind (2023) puts six modalities, image, text, audio, depth, thermal, and IMU, in one embedding space.

Video generation

Make-A-Video, September 2022, was Meta's first text-to-video system. Emu, announced September 2023, was Meta's first image-generation foundation model, trained on 1.1 billion image-text pairs. Emu Video and Emu Edit followed in November 2023. Movie Gen, announced October 4, 2024, is a 30B-parameter text-to-video model plus a 13B audio model for synchronized soundtracks, making up to 16-second 1080p clips. No weights were released and there is no public product. As of August 2026 it is still in a partnership phase with creative-industry feedback, and its capabilities have surfaced in Meta AI, Edits, and Vibes instead.

Audio

Voicebox, June 16, 2023, was the first generative speech model to generalize across tasks, but Meta declined to release the model or code over voice-impersonation risk. Audiobox, announced November 2023, unified speech, sound effects, and soundscapes with voice and natural-language prompts, and was released only to selected researchers. Its demo is no longer available as of February 2026. AudioCraft, open-sourced August 2, 2023, bundles MusicGen for text to music, AudioGen for text to sound, and EnCodec, a neural audio codec.

Language and translation

NLLB, or No Language Left Behind (2022), translates text across 200 languages. SeamlessM4T, August 22, 2023, is an all-in-one translation model covering speech and text in up to about 100 languages, under a non-commercial research license. MMS, or Massively Multilingual Speech (2023), covers speech recognition and text to speech in more than 1,100 languages. Chameleon, from May 2024, is a 7B and 34B early-fusion multimodal model trained from scratch on interleaved text and image tokens; its image generation was disabled in the open weights for safety, and its architecture is the blueprint for Llama 4's native multimodality.

Safety

Llama Guard 4, from late April 2025, is a 12B-parameter natively multimodal safety classifier, pruned from Llama 4 Scout, classifying prompts and responses against 14 hazard categories. Prompt Guard (2024, with a second version in 2025) is a tiny jailbreak and prompt-injection detector. LlamaFirewall (2025) is a guardrail framework for AI agents. All sit under Meta's Purple Llama umbrella.

Robotics

Meta Motivo, December 2024, is a behavioral foundation model for humanoid control, a single policy trained with unsupervised reinforcement learning and imitation of unlabeled motion-capture data, open on GitHub. Research, not a product, and cited as a building block for Meta's robotics ambitions.

Which one do I use?

Building agents on the Meta Model API

Ember for volume, Spark 1.3 for the everyday workhorse, Wonder when the task is long-horizon and worth the price. Meta's own pattern is to pair them: Wonder plans, Ember executes.

Self-hosting open weights

Spark Open if you want the strongest open Muse model and can serve it. Glimmer 30B if you want something that runs locally. Llama 3.3 70B if you want the proven dense workhorse, or Llama 4 Scout and Maverick if you want the newest open Llama generation, keeping in mind that hosted options for them are thinning.

Vision work

SAM 3.1 through the Model API for segmentation, detection, and tracking. DINOv3 as a frozen backbone for anything that needs dense visual features.

Everything else

The research models are mostly not things you can use. Movie Gen, Voicebox, and Audiobox never shipped publicly. Chameleon and the JEPA models are architecture, not products. If a Meta model is not on the Model API or Hugging Face, assume you cannot run it.

At a glance: the API-priced models