The Muse Glossary: Every Term This Site Uses, Defined

Muse is Meta's family of agent products: a personal agent called Muse, a closed developer model called Muse Spark, an open-weights model called Muse Glimmer, a terminal coding agent called Muse Code, and a connector system that links the agent to outside services. This glossary defines every term the catalog uses, in alphabetical order, with a source for anything factual. If a word shows up on a build page and you are not sure what it means, it should be here.

For the longer picture, read What is Muse first. For the timeline and the numbers, Start here has them.

Terms from A to C

Accounts Center. Meta's shared account settings. Facebook, Instagram and Threads connect to Muse automatically through it when those accounts are linked, which is why builds like the Instagram DM screener need no setup step (Meta help).

ACP (Agent Client Protocol). An editor-side protocol for driving coding agents. It is not a Meta standard, but builders have bridged Muse Code into it so it runs inside Zed and JetBrains IDEs; see muse-acp and muse-code-acp.

Apache 2.0. The license on Muse Glimmer's weights. It allows commercial use, modification and redistribution, including fine-tuned versions (model card).

Approval. Muse's habit of asking before important actions. By default it will not send an email or make a purchase without a yes from you, and some connectors can be held to read-only (Meta help).

BF16. Brain floating point, 16-bit. Glimmer's full-precision weights take about 58GB in BF16, per Unsloth. Most people run a quantized build instead.

Build. This site's word for any catalog entry: a repo, a post, a video, a demo, a benchmark. Every build links to its original source. Browse them by category or by type.

Cached input. Input tokens the Meta Model API has already seen as a repeated prefix. They cost $0.15 per million on the standard tier versus $1.25 for fresh input (pricing).

Category. One of eight topical shelves on this site: agents and browsers, games and real time, triage and routing, trading and markets, content and growth, research and data, robotics and devices, and tools and apps.

Contributor model. A cheaper tier of Muse Spark on the Meta Model API (muse-spark-1.3-contributor, muse-spark-1.2-contributor). It costs $0.10 per million input tokens and $0.20 output, and Meta may train on the data you send (pricing). Muse Code uses it by default, according to DataCamp.

Credential surrogation. Muse's way of keeping secrets away from the model. The agent's code only handles a stand-in token, and a separate authority swaps in the real one at the network boundary (Meta research).

Custom connector. A connector Muse writes for one user from a service's API. Credentials go into a secure store. Meta does not review custom connectors (Meta help). The Linear connector built in under a minute is a typical example.

Terms from D to L

DFlash. A speculative decoding method shipped alongside Glimmer. Meta's model card reports 3.1x on an RTX 5090 (74.9 to 233.4 tok/s) and 1.8x on an M5 Max. Builders report mixed results; one M5 Pro test measured about 1% speedup.

Directory connector. A connector listed in Muse's built-in directory after Meta's functional, security and legal review plus end-to-end testing (muse.ai/platform).

Effort level (reasoning strength). A dial for how long a model thinks. Glimmer exposes low, medium, high and xhigh. Spark 1.3 adds a max mode. Several builds, like VulcanBench, test every level.

Everyday, High and Power (Muse Code plans). Muse Code's subscriptions: $5, $15 (5x usage) and $50 (20x usage) per month (dev.meta.ai). Not to be confused with the Muse agent's Power plan.

GGUF. The file format used by llama.cpp, LM Studio and Ollama for quantized models. Unsloth lists Glimmer GGUFs from 2-bit to 8-bit.

Glimmer. Short for Muse Glimmer 30B, Meta's open-weights model, released August 10, 2026. About 29.6B parameters, dense, 131,072-token context, text and image in, text out (model card). It is self-hosted, not served by the Meta Model API.

Goals. A Muse agent feature for setting objectives the agent works toward; How I AI walks through goal setting in its review. The clearest outside write-up is a Codex user's critique of how Goals are designed.

KV cache. Memory a model keeps for each token of context. Glimmer uses about 52 KiB per token thanks to 2 KV heads and 3:1 sliding-window attention (Sebastian Raschka), which is part of why long contexts fit on consumer GPUs.

Link (Stripe Link). The payment wallet Muse uses. At 1M+ Link merchants Muse checks out with your saved method; elsewhere Link issues a single-use virtual card for that purchase (Stripe).

Terms from M

Max (Muse plan). The $100/month Muse subscription with 3B Muse tokens per week (Meta help). TechCrunch calls it "Maximum".

Max reasoning mode. Muse Spark 1.3's highest effort setting. Meta's OSWorld 2.0 number (66.9) is at max.

MCP (Model Context Protocol). An open protocol for exposing tools to AI models. Muse custom connectors can talk to remote MCP servers over streamable HTTP, according to Parallel. See MCP servers vs Muse connectors.

Meta Model API. Meta's hosted API for Muse Spark and related models, at https://api.meta.ai/v1. It accepts OpenAI Responses, Chat Completions and Anthropic Messages formats (docs).

MLX. Apple's machine learning framework for Apple Silicon. Glimmer runs through it via ports like muse-glimmer-mlx and an Ollama 30b-mlx tag.

Muse. The personal agent itself, launched September 8, 2026 for US adults. It runs errands, fills forms, books travel and buys things, and keeps working after you close the app. It runs on Muse Spark 1.3 (Meta).

Muse Charm. A pocket device with a real-time voice model, announced at Connect 2026. Details are due "later this year" (Meta).

Muse Code. Meta's terminal coding agent, co-trained with Muse Spark. It runs subagents in isolated git worktrees and keeps replayable event logs (Meta).

muse exec. A Muse Code command for handing it a single task non-interactively. Builds such as shunt-muse and delegate-to-muse use it to route work from other agents.

Muse for Mac. The macOS app, released September 17–18. It can work with iMessage, Notes, Reminders, Mail and Calendar, each set to Off, Read only, or Read and interact (Meta help).

Muse Image. Meta's image model on the Model API (muse-image-1.0), $0.01 per image.

Muse Realtime Avatar. An animated avatar for the agent, announced at Connect 2026.

Muse Secure VM. The dedicated virtual machine each Muse user gets, built on systemd-nspawn runtime cells. One builder reports about 2 vCPUs, 8GB RAM and 100GB disk.

Muse Session Protocol (MSP). The protocol Muse Code exposes for driving sessions, used by Helicon via muse serve and bridged to ACP by muse-acp.

Muse Spark. Meta's closed-weights reasoning model line. Version 1.3 (September 2) has a 1,048,576-token context and is tuned for long-horizon coding and agent work (Meta).

Muse tokens. The unit of Muse agent quotas. Free is usage-limited (The Batch reports up to 100M per week), Power is 500M per week, Max is 3B. Meta has published no conversion from tokens to tasks.

Muse Voice Transcribe. Speech-to-text on the Model API (muse-voice-transcribe-1.0), $0.18 per audio hour.

Terms from N to R

NVFP4. NVIDIA's 4-bit floating-point format for Blackwell GPUs. Several DGX Spark builds, like this tuning repo, serve Glimmer in NVFP4.

Open weights. Model weights you can download and run yourself. Glimmer is open weights; Spark is not, though Meta says a "Muse Spark open weights release" is coming.

OpenAI-compatible. An API that accepts requests shaped like OpenAI's. The Meta Model API is, and so are most local Glimmer servers, which is why so many existing tools work with a base URL change.

Parallel tool calls. Several tool calls in one model turn. The Model API allows them by default (parallel_tool_calls: true); see structured outputs and tool calling.

Power (Muse plan). The $20/month Muse subscription with 500M Muse tokens per week.

Prompt caching. Reusing a repeated prompt prefix to cut cost and latency on the Model API.

Prompt injection. Hidden instructions in web pages, emails or files that try to steer an agent. Muse layers model training, labeling of untrusted input, classifiers and human approval against it, and Meta's bug bounty pays up to $130,000 for a single-user prompt injection.

Quantization. Storing weights at lower precision to save memory. Glimmer's 4-bit build is about 17GB per Unsloth. See running Glimmer on a consumer GPU.

Read only / Read and interact. Per-app permission levels in Muse for Mac.

Terms from S to Z

Sentinel. The separate authority in Muse that gates connector actions and network egress, and swaps surrogate tokens for real credentials (Meta research).

Skill. Saved instructions an agent reuses. In Muse, a connector is the code and credentials; a skill is the instructions for using them (Parallel). In Muse Code, skills include /plan, /grilling and /taste. On this site, Skills is also a build type.

Speculative decoding. Using a small draft model to propose tokens a big model then verifies. DFlash is Glimmer's version.

Structured output. Constraining a model's reply to a JSON schema. The Model API supports it through response_format or text.format.

Tool search. A Model API feature where tools marked defer_loading: true load only when the model asks for them, saving tokens.

Type. This site's format label: X posts, Reddit posts, GitHub, videos, sites, skills, resources and guides.

Unsloth. A fine-tuning and quantization project with day-one Glimmer support, free notebooks and GGUF builds. See fine-tuning Glimmer.

vLLM and SGLang. Server-side inference engines for serving models to many users. Both had day-0 Glimmer support. SGLang reports about 230 tok/s on one RTX 5090 with NVFP4 and DFlash (SGLang).

Worktree subagents. Muse Code runs subagents in isolated git worktrees, so parallel tasks do not trample each other's files.

If a term is missing, submit a build that uses it and we will add it here.

Frequently asked questions

What is the difference between Muse and Muse Spark?

Muse is the consumer personal agent you talk to in the app, on the web or in WhatsApp. Muse Spark is the model family underneath it, and version 1.3 powers the agent. Developers can call Spark directly through the Meta Model API.

Is Muse Glimmer the same as Muse Spark?

No. Glimmer is a separate 30B open-weights model distilled from Muse Spark and released under Apache 2.0. You run it yourself on your own hardware or a third-party host, while Spark is closed and served by Meta.

What does "contributor" mean in Muse Spark model names?

Contributor models are a low-price tier on the Meta Model API where Meta may train on the data you send. They cost $0.10 per million input tokens and $0.20 per million output tokens, against $1.25 and $4.25 for the standard tier, which is not used for training.

What is a Muse token?

It is the unit Meta uses for Muse agent quotas. Power includes 500M per week and Max includes 3B per week. Meta has not published how many tokens a typical task uses, so the plans are hard to compare against real workloads.

What is the difference between a directory connector and a custom connector?

Directory connectors are reviewed by Meta and listed in Muse's built-in directory. Custom connectors are written by Muse for a single user from a service's API and are not reviewed by Meta.