What Muse Is
A plain-language tour of Meta's Muse family: why a company that builds both the model and the coding agent matters, what Muse Spark and Muse Glimmer are for, how you reach them, and what they cost.
What Muse Is
A plain-language tour of Meta's Muse family: why a company that builds both the model and the coding agent matters, what Muse Spark and Muse Glimmer are for, how you reach them, and what they cost.
You come away able to say what a harness is, name the two models the film covers and what each is for, see how the standard and contributor price tiers differ and what the cheaper one costs you, and know which numbers to confirm yourself before you rely on them.
Honest about how it was made: this film was drafted in Muse and then cleaned up and rebuilt with Claude. Where the film departs from the course it began from, it follows Meta's own pages: Spark 1.3 is the latest version (1.2 the previous one), Spark takes text, images, video, audio and PDFs in but answers in text only, Glimmer is released as open weights under Apache 2.0, and the prices are $1.25 in and $4.25 out per million tokens on the standard tier against $0.10 and $0.20 on the contributor tier, with 3,000 versus about 100 requests a minute.
A note that opens the film Muse drafted this, Claude cleaned it up. This film began as a rough draft made with Muse. Muse runs on a simple model called Spark, and Spark makes a lot of mistakes, so Claude cleaned the draft up. The method is the point: draft in Muse, then finish in Claude or Codex. One kind of mistake we looked for: a draft Muse diagram of two big boxes and a check mark under a question about whether a non-technical person could use it, with both boxes empty. A diagram that shows nothing teaches nothing, so Claude replaced diagrams like this one.
Narrated by Liam, in for Bear. Narration is an AI voice; the film was made with AI tools.
Chapters 0:00 Muse note 0:28 The idea and the key terms 1:00 The Bet 2:01 The Models 3:13 Reaching The Models 3:38 The Price Tag 4:13 The Catches 4:40 The Honest Picture 5:07 Recap 5:37 Your Turn
Your turn Paste this into Claude: I'm deciding how to use Meta's Muse models for [describe your project in two sentences]. Compare four options for me: Muse Spark on the standard tier, Muse Spark on the contributor tier, Muse Glimmer running on my own hardware, and a model I already use. For each, list what drives the cost, what data I give up, and the main risk. Don't recommend one yet. Then check two things yourself. Does it say the contributor tier trains on your prompts? And does it name a price you would still need to confirm on Meta's pricing page?
Sources shown on screen Meta's public developer pages at dev.meta.ai (models, pricing and rate limits, Muse Glimmer, coding agents, SDKs, overview, the Muse Code product page and its launch post), captured 2026-10-06 and 2026-10-07. Narration follows a Muse course taught by Andrew Brown, named on screen. Nothing here is a product endorsement.
Part of the playlist Muse for Educational AI.
@NikBearBrown
What this video covers
- 0:00 — Muse note
- 0:28 — The idea and the key terms
- 1:00 — The Bet
- 2:01 — The Models
- 3:13 — Reaching The Models
- 3:38 — The Price Tag
- 4:13 — The Catches
- 4:40 — The Honest Picture
- 5:07 — Recap
- 5:37 — Your Turn
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
Muse note
[0:00] A note before we start, this film began as a rough draft made with Muse. Muse runs on a simple model called Spark and Spark makes a lot of mistakes. So, Claude cleaned this draft up. The method is the point, draft in Muse then finish in Claude or Codex. Here is the kind of mistake we look for. A draft diagram from Muse, two big boxes and a check mark under a question about whether a non-technical person could use it. Both boxes are empty. A diagram that shows nothing teaches nothing. So, Claude replaced diagrams like this one.
The idea and the key terms
[0:28] Hello, this is Liam and Forbear. You may know Muse as AI app. This lecture is about what sits underneath, a family of models and a coding agent that Meta built to run them. What the models are, how you connect to them, and what they cost. Five terms, a harness, the program around a model that reads files, runs commands, and edits code. Muse Spark, Meta's hosted model, which you call over the internet. Muse Glimmer, an open model you download and run yourself. Open weights, the train numbers are published so anyone can run the model. And a token, a small chunk of text, the unit you pay by.
The Bet
[1:00] A coding tool has two halves. The model is the trained brain. The harness is the program wrapped around it. It reads your files, runs commands, and edits code. The halves are separate pieces. You can point open code, Claude code, or Codex at the same model and each one will drive it. Muse Spark works inside all three. Meta builds both halves. Spark is the model and Muse code, Meta's own coding agent, is the harness. One company owning the whole stack is called vertical integration. Andrew Brown, who teaches a course on Muse, argues that is an option worth using. Here is Muse code on Meta's own site, Meta's first coding agent. One command installs it in your terminal. Meta's pitch for the pairing sits on the same page. Code trained for tight integration with better tool use and fewer retries than a generic wrapper. Meta's launch post says how. Muse code was in the training loop from day one, so Spark learned to work with its tools. Meta also trained across other harnesses, so Spark still carries over to the coding agents you already use.
The Models
[2:01] Meta ships five model families. This lecture covers two, Spark and Glimmer. The others make images, transcribe speech, and outline objects in photos. And Llama, the famous sibling, gets this single line. Spark is the managed model. You call it over the internet and Meta runs the hardware. It comes in numbered versions, 1.1, 1.2, and now 1.3. The course teaches on 1.2. Spark is multimodal. It reads text, images, video, audio, and PDFs, and it answers in text. Pictures come from a separate model, Muse image. Glimmer is the open model. It has 30 billion parameters, the numbers it learned, and Meta trained it on Spark's own outputs. It saw more than 100 languages, and the weights are released under the Apache 2 license. And it fits on one graphics card. The full precision weights are about 60 GB, too big for a 32 GB card like the RTX 5090. Meta's compressed build is under 17 and runs in about 19 GB. How good are they? Meta's own launch charts give a first answer. Spark 1.2 is second on Terminal Bench and third on DeepSWE. Andrew Brown calls that Goldilocks, upper middle on most leaderboards, third best at everything.
Reaching The Models
[3:14] Reaching Spark is a plain web request. Meta offers three request formats. Responses and chat completions work like Open AI's. Messages works like Anthropic's, which is how Claude code connects. So, your existing code barely changes. The docs say there's no custom client to learn. Point the Open AI SDK at Meta's base URL, add your key, and keep the rest. The Anthropic SDK can be pointed at Meta, too.
The Price Tag
[3:39] Now, the price tag. For the API, there's no subscription. You pay per token, a small chunk of text for what goes in and for what comes out. Use a little, pay a little. Here is Meta's pricing page. The standard tier costs $1.25 per million tokens in and $4.25 per million out. Every Spark version shares that price, and Meta says your data is not used for training. The cheaper option is the contributor tier. Input falls to 10 cents per million tokens and output to 20 cents. The trade, you give Meta permission to train on your prompts and the model's answers.
The Catches
[4:13] There is a second catch. The contributor tier allows about 100 requests a minute. The standard tier allows 3,000. Cheap and slow or standard and fast. Two extras, searching the web adds $2.50 per thousand queries on top of the tokens. And the course notes there is no batch discount. Meta's pricing page lists none. One wrinkle, Muse Code, the coding agent, sells monthly plans of its own. That is separate from paying per token on the API.
The Honest Picture
[4:40] First, the honest limit. In Nick's experience, Spark is noticeably less capable than the very top models. Meta's own chart agrees it is not first. Don't pick it for the leaderboard. Pick it for the price, for the open Glimmer you can run yourself, and for the tools it already works with. That is the case for Muse. One last thing, versions, prices, and limits change. Every number here came from Meta's own pages, captured on October 6th, 2026. Check them before you
Recap
[5:07] Build. Let's recap with Claude. Meta builds both halves, the model and the harness. Spark is the managed model, and Glimmer is the open one that fits a single graphics card. The API speaks the Open AI and Anthropic formats, so your code barely changes. Standard costs $1.25 in and $4.25 out per million tokens, and contributor costs a dime and 20 cents. The cheap tier is rate limited. Search costs extra. And Muse Code has plans of its own. And Spark is not the top model, so pick it for price, not the leaderboard.
Your Turn
[5:37] Your turn. Paste this into Claude. I'm deciding how to use Meta's Muse models for describe your project in two sentences. Compare four options for me. Muse Spark on the standard tier, Muse Spark on the contributor tier, Muse Glimmer running on my own hardware in a model I already use. For each, list what drives the cost, what data I give up, and the main risk. Don't recommend one yet. Then check two things yourself. Does it say the contributor tier trains on your prompts? And does it name a price you would still need to confirm on Meta's pricing page? What Muse is at Nick Bear Brown.
