Muse Wonder Full
An AI model released by Meta Superintelligence Labs and part of the Muse series. It is a more complete version of Muse Wonder, offering greater capability or a larger parameter scale.
Core Facts
Muse Wonder is offered through the Meta Model API and Muse Code; Muse Wonder Full is for Muse Code subscribers.
Meta describes two model features: Active Memory carries relevant context across conversations, and Active Focus keeps trying different paths on difficult goals and confirms before critical actions.
In coding agent scenarios, Claude Code on Opus 5.5 scored 66, Wonder scored 63, and only Wonder Full can reach 68.
Timeline (last 90 days)
Meta's Muse Wonder Full performed notably weakly in the 3D rendering test, and Meta AI head Alexander Wang publicly admitted, "we forgot to make it good at 3D modeling."
Wonder Full scored 57.5 on the SWE-Marathon benchmark and reached 79.0 on DeepSWE 1.1.
Meta reported experimental results stating that Wonder Full's performance on unfamiliar tasks improved relative to the corresponding trend of the previous-generation model, which Meta attributes to a stronger base model.
Meta divides test tasks into In-distribution tasks, which were a post-training optimization focus, and Out-of-distribution tasks, which received no targeted optimization, and compares how Muse Spark 1.3 and Muse Wonder Full perform under different reasoning budgets.
Muse Wonder Full's 57.5 score on SWE-Marathon used the xhigh reasoning setting.
The lava lamp test targeted the regular version of Muse Wonder, while some of the standout results Meta published came from Wonder Full, and there is a capability gap between the two.
Muse Wonder Full scored 30 on the AA-Omniscience factual recall test.
Knowledge is a weak spot for Muse Wonder Full, whose AA-Omniscience score is lower than that of Claude Opus 5.5.
Setting Wonder's reasoning effort to max does not make it become Full.
In coding agent scenarios, Claude Code on Opus 5.5 scored 66, Wonder scored 63, and only Wonder Full can reach 68.
40 more timeline events
All Facts (20)
Muse Wonder is offered through the Meta Model API and Muse Code; Muse Wonder Full is for Muse Code subscribers.
80%VerifiedMeta describes two model features: Active Memory carries relevant context across conversations, and Active Focus keeps trying different paths on difficult goals and confirms before critical actions.
65%Partially VerifiedIn coding agent scenarios, Claude Code on Opus 5.5 scored 66, Wonder scored 63, and only Wonder Full can reach 68.
70%UnverifiedClaude Opus 5.5 is the absolute flagship of Anthropic's product line, while Muse Wonder is positioned as the smaller of Meta's two models, with its larger version being Muse Wonder Full (available only via Muse Code).
60%UnverifiedMuse Wonder Full scored 56 on the Artificial Analysis v4.3 leaderboard, second globally, behind Claude Opus 5.5's 58; the standard version of Muse Wonder scored 52.
60%UnverifiedThe lava lamp test targeted the regular version of Muse Wonder, while some of the standout results Meta published came from Wonder Full, and there is a capability gap between the two.
50%UnverifiedMeta divides test tasks into In-distribution tasks, which were a post-training optimization focus, and Out-of-distribution tasks, which received no targeted optimization, and compares how Muse Spark 1.3 and Muse Wonder Full perform under different reasoning budgets.
50%UnverifiedMeta reported experimental results stating that Wonder Full's performance on unfamiliar tasks improved relative to the corresponding trend of the previous-generation model, which Meta attributes to a stronger base model.
50%UnverifiedWonder Full scored 57.5 on the SWE-Marathon benchmark and reached 79.0 on DeepSWE 1.1.
50%UnverifiedMuse Wonder Full's 57.5 score on SWE-Marathon used the xhigh reasoning setting.
50%UnverifiedMuse Wonder Full scored 30 on the AA-Omniscience factual recall test.
50%UnverifiedKnowledge is a weak spot for Muse Wonder Full, whose AA-Omniscience score is lower than that of Claude Opus 5.5.
50%UnverifiedSetting Wonder's reasoning effort to max does not make it become Full.
50%UnverifiedSome users reported that a single Wonder Full run consumed 38% of the weekly Muse Code quota, indicating that long-duration agent tasks are extremely compute-intensive.
50%UnverifiedMeta released two models: Muse Wonder and Muse Wonder Full, but only Wonder (muse-wonder-1.0) is available on the API; Wonder Full is offered only through the paid Muse Code subscription plan, with no API access.
50%UnverifiedWhat actually beat competitors in the benchmarks is Wonder Full; most of the impressive results come from the Full version, not the Wonder version available via API.
50%UnverifiedMuse Wonder Full is aimed at Muse Code subscribers and is more complete in features, but the specific differences have not yet been explained in detail officially.
50%UnverifiedMuse Wonder is divided into two versions: the base version Muse Wonder (about 120 tps) and Muse Wonder Full (about 50 tps).
50%UnverifiedThe reviewer believes MuseWonder is suitable for users who need cross-application agent collaboration, document analysis and personalized planning, rather than developers pursuing ultimate coding quality.
50%UnverifiedThe Meta model card footnote explains: the score selection method prioritizes comparable third-party results, otherwise uses the highest of internal validation runs, self-reported scores, or official leaderboard scores; competitor runs are best-effort and not optimized for proprietary models; the SWE-Marathon comparison involves different agent harnesses; Wonder Full's 57.5 on SWE-Marathon is reported with xhigh reasoning effort.
50%