Latest2 Sep 2026Reasoning LLM
Muse Spark 1.3
Muse Spark 1.3 is the latest version of Meta Superintelligence Labs' Muse Spark model, tuned for long-running agent work and coding. Meta says it finishes coding tasks with about 20% fewer tool calls and 25% fewer tokens than version 1.2. It is the model family behind Meta's Muse personal agent, launched six days later.
- Longer-horizon agent work across several workflows in one thread
- About 20% fewer tool calls and 25% fewer tokens than 1.2 (Meta)
- Better instruction following on complex multi-step tasks
- Improved resistance to adversarial inputs and prompt injection (Meta)
AccessMeta Model API and Muse Code (dev.meta.ai); powers the Muse personal agent; closed weights
1 Sep 2026Audio/voice model
Muse Voice Transcribe
Muse Voice Transcribe is Meta Superintelligence Labs' first real-time audio perception model. It handles streaming speech recognition, speaker identification and end-of-speech detection in a single model, for live transcription and voice agents. Meta reports a 3.1% streaming word error rate and support for more than 70 languages, 25 of them extensively verified.
- Streaming speech-to-text in 70+ languages (25 extensively verified)
- Speaker diarization for 20+ speakers
- Handles audio longer than one hour without post-processing
- Native code-switching between languages
- Streaming word error rate 3.1% (Meta)
AccessMeta Model API, Meta AI for Mac and Muse Code
10 Aug 2026Open-weights LLM
Muse Glimmer
Muse Glimmer is a 30-billion-parameter open-weights model from Meta, built to run agent workflows locally on a single consumer GPU, even without an internet connection. It was distilled from Muse Spark and accepts text and images. Released under the Apache 2.0 licence, it is the first open-weights model in the Muse family.
- 30B parameters; text and image input
- Quantised to about 4-bit, under 20 GB, for consumer GPUs
- Agentic tool calling and multi-step reasoning
- Works offline on a personal computer
AccessOpen weights on Hugging Face (meta-models/Muse-Glimmer-30B), Apache 2.0 licence; runs locally, including offline
5 Aug 2026Coding model
Muse Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, released together with Muse Code, Meta's terminal-based coding agent. Meta says the model was co-trained with Muse Code and improves code generation, debugging, codebase understanding and end-to-end developer workflows. API access was widened to more countries at the same time.
- Code generation, complex debugging and codebase understanding
- Co-trained with the Muse Code terminal agent
- Muse Code runs async background agents with crash replay
- Evaluated on Terminal-Bench 2.1 and DeepSWE 1.1 (charts only)
AccessMuse Code (Meta's terminal coding agent) and Meta Model API, with expanded global access; closed weights
9 Jul 2026Reasoning LLM
Muse Spark 1.1
Muse Spark 1.1 upgraded Meta's first Muse model for agent tasks and opened the Meta Model API in public preview, giving developers broad access to a Muse model for the first time. Meta reports gains in tool and computer use, coding and multimodal understanding, and a 1-million-token context window.
- 1M-token context window
- Improved computer use across multiple applications
- Trained for multi-agent orchestration
- Stronger multimodal reasoning and perception
AccessMeta Model API (public preview); Thinking mode in the Meta AI app and meta.ai; closed weights
7 Jul 2026Image model
Muse Image
Muse Image is one of the first media-generation models from Meta Superintelligence Labs. It works as an agent: it calls search and coding tools to improve accuracy, refines its own images and uses more compute at generation time. It supports precise editing and combining several reference images, and works with Muse Spark to build GIFs, web pages and simple games.
- Text-to-image generation and precise image editing
- Composes from multiple reference images
- Calls search and code tools; self-refines its outputs
- No. 2 on Arena text-to-image and editing (Meta, 2026-07-05)
AccessMeta AI app, meta.ai, Instagram Stories (US), WhatsApp (some countries); Facebook coming; no API mentioned
8 Apr 2026Reasoning LLM
Muse Spark
Muse Spark is the first model from Meta Superintelligence Labs and now powers the Meta AI assistant. It is a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration. A 'Contemplating' mode runs several reasoning agents in parallel for hard problems, which Meta compares with Gemini Deep Think and GPT Pro.
- Text and image input; tool use and visual chain of thought
- Contemplating mode: parallel multi-agent reasoning
- Humanity's Last Exam 58% in Contemplating mode (Meta)
- Pre-trained with over 10x less compute than Llama 4 Maverick (Meta)
Accessmeta.ai and the Meta AI app, rolling out to WhatsApp, Instagram, Facebook, Messenger and AI glasses; private API preview for select users; weights not released
19 Nov 2025Embedding/other
SAM 3
SAM 3 is Meta's third Segment Anything model, a computer-vision model rather than an image generator. It finds, outlines and tracks objects in images and video from text prompts such as 'red baseball cap', example images or clicks. Meta released it with SAM 3D, two open models that rebuild 3D objects and human bodies from one image.
- Segmentation from text, example-image and click prompts
- Tracks objects through video
- Companion SAM 3D Objects and SAM 3D Body models
- SAM 3.1 update (2026-03-27) tracks 16 objects per pass
AccessOpen model weights and code (GitHub facebookresearch/sam3; Hugging Face); Segment Anything Playground for non-commercial trial