Muse Ember

Meta Superintelligence Labs' smallest and cheapest Muse model, released on November 9, 2026 as muse-ember-1.0 on the Meta Model API and in Muse Code. Meta describes it as "our smallest and most efficient model", a reasoning model for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing". It takes text, image, audio and video in, with a 1M-token context and up to 512K output tokens. Parameter count, architecture and training data are not disclosed, and no weights are released.

The one training fact Meta gives is lineage: when it launched Muse Wonder and Muse Wonder Full, it said it would use them "to develop specialized models", and "Muse Ember is the first trained using Muse Wonder Full". The post does not say whether that means distillation into new weights or post-training of an existing checkpoint, so this index records no provenance fields for it.

The pitch is cost, not capability. Pricing is $0.10 input, $0.004 cached input and $0.50 output per 1M tokens up to 200K-token prompts ($0.30 / $0.012 / $1.50 above that). Artificial Analysis scores it 36 on Intelligence Index v5 at max effort and 34 at xhigh, at $0.14 and $0.12 per index task. That matches Claude Haiku 5.5 (max) at about a third of its cost per task and sits one point under Muse Spark 1.3 (37, $2.39 per task), and 8 points under Muse Wonder (44). AA attributes most of the cost gap to cheaper cached input and roughly a third of Haiku 5.5's output tokens. Its weak spot in AA's testing is knowledge: AA-Omniscience accuracy is 21.1%, the lowest among similarly scored models, because it declines 64% of questions, which also keeps its hallucination rate at 17.4%.

Meta's self-reported results at max effort, against Claude Haiku 5.5 and GPT-6 Luna (both max): Terminal-Bench 4.0 40.9% (39.2 / 16.4), AA Coding Agent Index 46 in Muse Code (37 / 31), Agents' Last Exam 61.6% (57.0 / 57.1, Meta internal, Linux subset), OSWorld 2.1 69.5% partial / 33.8% binary, and PerceptionBench with tools 80.6%. It trails Haiku 5.5 on FrontierCode 2.0 (27.2% against 34.9%). Meta's chart gives AutomationBench-AA as 37.8%, while AA's own article gives 68.1% for the same configuration; the two sources disagree and this entry does not resolve it. Meta says Ember was evaluated under its Superintelligence Scaling Framework and stayed below the framework's high-risk thresholds for chemical and biological, cyber and loss-of-control risks. The license is proprietary.