Muse Wonder Full vs Standard Wonder: What KingBench 4 Actually Found
Meta's Muse Wonder Full tested against standard Wonder on six coding builds via KingBench 4, with scores, failures, and fixes checked.
What is Muse Wonder Full?
Muse Wonder Full is a variant of Meta’s Muse Wonder model family, a proprietary reasoning model line Meta announced on October 12th. Wonder supports a context window just over a million tokens, accepts multiple input types, and generates code through Muse Code, Meta’s coding tool that lets the model inspect projects, write files, run commands, and test its own output. The weights are closed, so there’s no local download option. Full is a distinct selection inside the native Muse CLI, not something you get by turning on maximum reasoning in the standard model. Picking “maximum reasoning” on standard Wonder keeps you on standard Wonder. You have to explicitly choose Muse Wonder Full to run it.
TL;DR
- Full scored higher than standard Wonder on three of six KingBench 4 builds (emulator, keyboard configurator, Pokémon-style game) but tied or scored lower on the rest, so the upgrade isn’t a blanket improvement.
- The headline benchmark gains are real but modest: Full lands fifth on Val’s index at 64.57% versus standard Wonder’s 61.06% in 10th, and scores higher on Terminal Bench 4 and Vibe Code Bench, though it actually loses to standard Wonder on code migration.
- Full is a subscription-only model available through Meta’s power usage ($50/month) and maximum usage ($200/month) tiers at launch, with no published pay-as-you-go API price, unlike standard Wonder which has both subscription and API access.
- Diagnostic passes don’t equal feature completeness: Full’s Game Boy emulator passed 23 of 29 public diagnostics, but DS and 3DS support stayed shallow, with missing GPU/audio services and a 3D state-saving setup that would balloon to roughly 14 gigabytes of stored history.
- Storage and integrity failures capped the Docker-alternative score at 3.0 out of 10, identical to standard Wonder, after an image import wrote outside its intended directory and an exported image failed its own hash check.
- The 3D room designer regressed slightly from standard Wonder (7.0 vs 7.6) despite a visibly nicer interface, because of unit-conversion bugs, incomplete autosave, and leaked graphics resources after repeated edits.
- The Pokémon-style game build held together as a continuous playable loop from the starting town through the first gym battle, suggesting Full handles multi-system integration tasks better than isolated technical builds.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How was Muse Wonder Full tested?
The evaluation used KingBench 4, a set of six coding build tasks: a multi-console emulator, a Docker alternative, a 3D interior design app, a keyboard configurator, a Pokémon-style game, and a Fitbit-style desktop app. Both Wonder runs (standard and Full) used the same prompts, the same Muse Code CLI version, high reasoning settings, two concurrent tasks, and fresh project folders, so the comparison isolates the model choice rather than tooling differences. Builds were allowed to run to completion with no added time limit or step budget. After each finished, the files were frozen and tested as a snapshot, with no manual repairs before scoring. Scoring weighed working requirements, correctness, presentation, and efficiency, with missing or unverified features counted against the result.
This matters because public benchmarks and hands-on build tests measure different things. A model can post a strong number on an abstract coding benchmark and still ship an application with storage boundary bugs or broken save states. KingBench 4 is designed to surface exactly those gaps by forcing a full application through independent functional checks after the fact.
How do the benchmark numbers compare?
On Val’s published index from the October 12th snapshot, Full ranks fifth at 64.57% against standard Wonder’s 61.06% in 10th place. That index blends several professional task categories, so it isn’t a direct measure of consumer app-building skill. On Terminal Bench 4, Full is listed fifth at 58.08%, tied visually with another model one slot above it. On Vibe Code Bench, Full scores 86% (12th place) compared to standard Wonder’s 83.09% (18th place), a clearer gap. But on code migration, Full’s estimated score (61.03%) actually trails standard Wonder’s 62.55%, showing the improvement isn’t uniform across task types.
Artificial Analysis gives Full an intelligence index score of 56 against 52 for standard Wonder, and its separate coding agents index (which scores the agent plus model together) puts Muse Code with Full at 68 versus 63 with standard Wonder. Those published configurations run at maximum effort, while the KingBench 4 builds ran at high reasoning, so the settings aren’t identical across sources. The takeaway: Full generally scores better on aggregate benchmarks, but the margin varies by task, and at least one benchmark shows it losing ground.
Where did Full actually beat standard Wonder?
The clearest wins came on the emulator, the keyboard configurator, and the Pokémon-style game.
On the emulator, Full’s Game Boy implementation passed 23 of 29 public diagnostics, including all 11 CPU instruction tests, and produced working battery saves and deterministic state restoration. Standard Wonder’s attempt was weaker overall. But DS and 3DS support stayed thin on both runs: ARM processors could run isolated test code but failed standard program counter and interrupt checks, and 3DS homebrew loading didn’t work. Full scored 3.1 out of 10 on this task versus standard Wonder’s 2.4, a real but incomplete improvement given the task asked for broad multi-console compatibility.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
On the keyboard configurator, Full’s “Hive Key” app scored 5.5 out of 10 against standard Wonder’s 3.5. Basic remapping worked against an independent firmware test fixture, and reconnect logic held up. But layer-switching keys generated wrong codes, encoder directions were reversed, and a macro meant to press Enter instead sent literal text because it wasn’t translated into the firmware’s macro format.
On the Pokémon-style game, Full delivered a playable loop from the opening sequence through catching a wild Pokémon, switching party members, healing, and reaching the first gym battle with working win and loss states. This kind of task rewards systems working together rather than any single deep technical feature, which seems to suit Full’s strengths better than isolated infrastructure builds.
Where did Full fall short?
The Docker alternative (“Lightbox”) tied standard Wonder at 3.0 out of 10. Despite implementing image operations, build commands, and lifecycle management, an independent test found that importing an image wrote through a symbolic link and altered a file outside the app’s intended storage directory, a real storage boundary failure. A separate bug meant images exported by the app failed their own integrity check on reload, because the saved content hash didn’t match the altered archive. No comparable performance benchmark against actual Docker workloads was available to back the app’s claimed efficiency advantage.
The 3D interior designer, “Atelier,” actually scored below standard Wonder (7.0 versus 7.6) despite a visually more polished interface with working box, cylinder, sphere, and cone creation, functional lighting controls, and real-time geometry editing. The regression came from usability and reliability defects: switching units to feet changed the label but not the underlying math, undo sometimes failed to revert changes, and autosave missed the user’s most recent edit in a reload test. Repeated editing also leaked graphics memory, with allocated textures and geometries climbing well past the actual active scene content, a sign of missing cleanup code.
Is Muse Wonder Full worth the upgrade?
Based on this six-task comparison, Full is a meaningful but inconsistent upgrade. It clearly improves results on tasks that reward broad system integration, like a playable game loop or a multi-protocol device configurator. It does not reliably improve, and in one case actively regresses, on tasks involving fine-grained state management, storage safety, or resource cleanup. Anyone evaluating Full for a specific coding workload should weigh which failure mode matters more: incomplete feature coverage (as with the emulator’s DS and 3DS support) or integrity bugs that corrupt saved data and exported files (as with the Docker alternative and the room designer’s autosave).
Pricing adds another consideration. Full is only available through Meta’s subscription tiers, listed at $50 and $200 a month at launch, with no published pay-as-you-go API rate. Standard Wonder offers both subscription and direct API access, which matters for anyone trying to estimate per-task costs rather than paying a flat monthly fee.
Frequently Asked Questions
What is KingBench 4?
KingBench 4 is a test set of six coding build tasks, including an emulator, a Docker alternative, a 3D design app, a keyboard configurator, a game, and a desktop utility app, used to evaluate how AI coding models handle complete, functional application builds rather than isolated code snippets.
Is Muse Wonder Full the same as turning on maximum reasoning in standard Wonder?
No. Full is a separate model variant that must be explicitly selected in the Muse CLI. Enabling maximum reasoning on standard Wonder does not switch the underlying model to Full.
Can I download and run Muse Wonder Full locally?
No. Meta has not released the weights publicly, so Full and standard Wonder are only accessible through Meta’s hosted subscription tiers or, for standard Wonder, its API.
Did Muse Wonder Full beat standard Wonder on every task?
No. It scored higher on three of the six KingBench 4 tasks (emulator, keyboard configurator, game), tied on the Docker alternative, and scored lower on the 3D interior designer.
How does Full’s pricing compare to standard Wonder’s?
Full is available only through subscription tiers listed at $50 and $200 a month at launch, with no published pay-as-you-go API rate. Standard Wonder offers both a subscription and direct API pricing, listed at $4 per million fresh input tokens and 20 cents per million cached input tokens.
