3D Generation

30 Models Every 5 Hours, Billed Against Your Coding Quota: Meta Puts Textured 3D Generation Inside Muse Code — and Says Nothing About Polycount or File Formats

November 21, 20260 views阅读中文原文
30 Models Every 5 Hours, Billed Against Your Coding Quota: Meta Puts Textured 3D Generation Inside Muse Code — and Says Nothing About Polycount or File Formats

TL;DR

On November 20, Meta published a short post on its developer site announcing that Muse Code can now generate 3D models. Muse Code is Meta's coding agent that runs in the terminal; until now it wrote code and called image models. From that day it also has a tool call: hand it a text description or a single image, and it produces a textured 3D mesh.

Three things worth remembering:

  • The capability ships to every Muse Code subscription tier. It is not a separate product with its own price tag. Usage is deducted from the quota you already pay for.
  • There are two model settings: Nano (fast and cheap, the default on the Everyday Usage plan) and Pro (higher quality, reserved for higher tiers).
  • The only quantified number Meta gave: the Everyday Usage plan can produce roughly 30 3D models within a 5-hour window.

It is equally important to say what this is not. It is not a new 3D generation model launch, and it is not a standalone website. It is a new capability bolted onto a coding agent, which means Meta still has not answered the question of whether it has its own 3D generation model.

This article is compiled from Meta's official blog post and from coverage of the announcement by multiple media outlets. Where Meta did not publish a number, we do not invent one.

What was actually announced

Muse Code is Meta's terminal coding agent, available on macOS, Linux and Windows, running on Meta's own Muse model family (Muse Wonder, Muse Spark, Muse Ember). Point it at a repository and it plans, edits files and runs commands on its own.

Here is how Meta describes the addition: Muse Code can now create textured 3D models through a tool call, in the same way it already calls Muse Image to produce diffusion-based images. Meta lists three intended uses: building games, sending models to a 3D printer, and visualising scenes.

The design decision that matters: there is no new entry point and no new interface. You do not open another website, create another account, or learn another parameter panel. You say "build a model for this character" in a normal instruction, and the agent decides whether to reach for the 3D tool.

Meta is explicit about one more thing: Muse Wonder, Muse Spark and Muse Ember can all call the 3D tool, and they will use it "when they seem like the best way to complete your tasks." In other words, the tool is not something you toggle from a menu. The model decides.

Why it is worth a look

The first layer of value is that 3D generation now sits inside the code-writing pipeline. Game prototyping used to be a fragmented loop: generate a model on site A, download it, rename it, drop it into the project, discover the scale is wrong, go back to site A and redo it. That step and the code now live in the same conversation, and the agent knows what size and orientation the project needs because it is reading your code at the same time.

The second layer is how it is billed. 3D generators usually sell credits per generation, a few cents to a few tenths of a dollar each, which forces you to do mental accounting on every attempt. Meta folded this into the subscription quota — for anyone already paying for Muse Code, the marginal cost is zero and the only question is whether the quota lasts.

The third layer is the real point: Meta has turned "generate 3D" into an ordinary agent tool. Tool-shaped means composable — build a model, write the loader code, spin up a web preview, then revise the model based on what the preview shows. Meta's two sample projects follow exactly that path: a playable 3D dungeon crawler with enemies, traps, spells and attacks, and a website that places different apple varieties side by side in 3D. Both download as zip files.

What Meta disclosed, and what it withheld

Disclosed:

  • Two model settings. Nano is fast and cheap; Pro is higher quality. The Everyday Usage plan only gets Nano. Higher tiers get Pro.
  • The quota figure. On the Everyday Usage plan, roughly 30 3D models per 5-hour window. This is the only concrete throughput number in the post.
  • Sample count. Meta published 25 sample models. Fifteen objects: castle, skyscraper, cathedral, cabin, iceberg, red rock formation, coral reef, gargoyle, bronze statue, fruit cart, bulldozer, tank, motorcycle, vintage pickup, battleship. Ten characters: tuxedo gent, cardigan guy, skeleton in a tutu, parka guy, cyborg soldier, chibi vampire, chibi assassin, mailbox monster, blue orc, goblin. Meta notes that all 25 were generated by the Pro model from a single input image.
  • One honest gap. There is no built-in rigging or animation tool. Meta does not pretend otherwise — it hands you a prompt to give Muse Wonder Full, instructing it to rig a character with a Mesh2Motion human skeleton and Blender automatic weights, and to deliver a web viewer you can use to inspect the character with simple animations.

Withheld, and these are the ones that bite in practice:

  • Polycount budget. This is the first gate on any 3D asset, and games and printing want opposite things. Meta gives no number at all.
  • Export file format. No word on whether you get GLB, OBJ, STL or FBX. For 3D printing this is life or death: STL carries neither textures nor units, GLB carries textures but not every slicer accepts it.
  • Watertightness and wall thickness. Meta says the output can be sent to a 3D printer, but does not promise watertight geometry, gives no minimum feature size, and does not say whether it targets FDM or resin. "Printable" is not the same standard for those two.
  • Per-generation time. The only throughput figure is 30 per 5 hours, which back-of-envelope works out to about 10 minutes each — but that includes queueing and your own conversation turns, so it is not generation speed.
  • Price. The 3D announcement itself contains no monthly figure.

Who it is a good fit for

  • You are prototyping a game and need a lot of placeholder assets. Ten goblin variants, twenty crates, a batch of building blocks. Generating them with Nano beats digging through asset libraries.
  • You need to show someone "something goes here." Scene visualisation benefits most: no modelling, just put the concept into the frame with one sentence.
  • You are making custom small objects for 3D printing. The official samples include tanks, pickups and battleships — hard-surface objects with clear silhouettes, which suggests that is where it is steadiest.
  • You already pay for Muse Code. In that case 3D is a freebie and the cost of experimenting is only quota.

Not a good fit: who should skip it

  • You need a rigged character that goes straight into an engine. No built-in rigging; skeletons and weights mean Blender, and that manual work is not smaller than the modelling itself.
  • You need engineering parts with exact dimensions. No unit convention, no tolerances, no dimensional parameters. Fit-critical parts are out.
  • You need a structural part to print on an FDM machine. No watertightness or wall-thickness guarantees, so leaks and thin walls are your risk to carry.
  • You need product-grade assets with PBR materials and 4K textures. Texture specifications are not mentioned at all, and the 25 samples only show appearance, not texture resolution.
  • You want the best 3D generation model available. This capability is positioned as convenient, not as best-in-class. On quality it has to be compared against Tripo, Meshy and Hunyuan, products that have been running on this track for years.

How to use it: a path you can finish today

Muse Code is a terminal tool, so the loop is short:

  1. Install. One command on macOS or Linux: curl -fsSL https://dev.meta.ai/install.sh | bash. There is a PowerShell installer for Windows. Sign in and attach your subscription.
  2. Enter a project. cd into your repository and start Muse Code. Try an empty directory first rather than letting it create files in a real project.
  3. Ask for a plan before it acts. Do not open with "make a castle." Open with "I need three props for this game scene; tell me how you plan to do it." The agent will lay out an approach, and you can redirect before anything is written.
  4. State the purpose when you ask for 3D. Add a line such as "this model is for 3D printing, keep it a closed solid, walls no thinner than 2mm," or "this model goes into Unity, watch orientation and scale." Different purposes lead to different prompting strategies — this is the one advantage an agent has over a pure generator, so do not waste it.
  5. Have it build the preview too. Meta's own samples work this way: the model plus a viewer you can orbit in a browser. You catch scale and orientation problems immediately instead of after an engine import.
  6. For rigging, copy Meta's prompt. Mesh2Motion skeleton, Blender automatic weights, plus a web viewer with simple animations.
  7. Specify the format before export. Since Meta does not state one, ask for it explicitly: "export GLB with embedded textures," or "export STL and tell me the units."

Pitfalls: four things that will bite

  • The quota is a rolling 5-hour window, not a monthly total. The "roughly 30" on Everyday Usage means 30 per 5 hours. Spend an afternoon generating models and you will have no quota left for writing code that evening, because 3D and coding draw from the same pool. This is the easiest trap to fall into.
  • Meta did not say how much quota Pro consumes. You know higher tiers "can make more" and "get Pro," but not how many Nano models one Pro model equals. Budget on the worst case.
  • The gap between Nano and Pro is not quantified. All 25 samples were made with Pro, which means the quality Meta is showing is invisible to the Everyday plan, which only gets Nano. This is the least transparent part of the post: what the cheapest tier actually buys you is unanswered.
  • Do not demand rigging. The tool generates characters, but the output is a static mesh. Making a goblin walk means Blender, and there is no shortcut.

One small thing that says a lot about the state of the launch: the official post returns a 500 to command-line fetches, and the body only renders inside a browser. The page was still being adjusted on launch day, which in practice usually means documentation and edge cases are still catching up.

Cost and time

Official position: 3D generation is charged against your subscription quota, with no separate fee. Everyday Usage gets roughly 30 models per 5-hour window.

Third-party tier list (not part of this announcement, provided only as context — the official 3D post names no price): Muse Code monthly plans are Everyday $5, High $15, Power $50 and Maximum $200, with usage measured per 5-hour window; the $15 tier is about 5× the $5 tier, $50 about 20×, and $200 about 80×. Muse Wonder Full is available only on Power and above. Those figures were checked by third parties in October 2026, and Meta's own threads account mentioned the four-tier structure in November, but Meta has not published how 3D usage converts into quota.

Rough arithmetic on that basis: 30 models per 5 hours on the $5 tier works out to about $0.17 of nominal quota per model. That is not "a model costs 17 cents," because you are also spending that quota on code. The real calculation depends on how much coding you still have to do this month.

On time, do not expect seconds. Thirty models per 5 hours averages out to about ten minutes each, conversation turns included.

Next step

Today: spend ten minutes in an empty directory running one pass of "one sentence in, one textured model out, orbit it in a browser." That tells you immediately whether Nano's quality is enough for your work, instead of guessing from someone else's screenshots.

This week: take one real, small need you actually have — a prop, a cover element, a small printable part — and carry it all the way into the software you really use. Wherever it gets stuck is the honest boundary of the capability.

Not yet: do not move a production pipeline onto this. Meta published no polycount, no format and no watertightness figure, so switching now means absorbing all the risk yourself. Wait for the documentation, or for someone to measure export formats and polycounts.

In one sentence: this is a release that makes 3D generation an ordinary agent tool, and the direction matters more than the parameters — but the numbers Meta gave are thin, so whether it works for you is a ten-minute test you have to run yourself.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...