Atlas

Models

Meta

Muse Ember

163.8
Capability · 95% interval 161.9–165.8
Rank 20 of 893 (interval 17–27)
Relative cost 0.067× · $0.0034 per short-answer task · $0.086 per agentic task

Muse Ember is Meta's smallest and most efficient Muse model, released November 9, 2026 for the work agents repeat: tool calls, coding subtasks, routing, extraction and long-document parsing. It is the first model Meta trained using Muse Wonder Full. It takes text, image, audio and video input with text output, and Meta gives a 1M-token context window, without an exact limit, and up to 512K output tokens. It is available through the Meta Model API and Muse Code. The API accepts minimal through max reasoning effort, with max the default. Parameter counts are undisclosed.

Launch post

Scored on 161 benchmarks · 176 total scores · 174 independently reported

Score providers

Organizations reporting headline scores for 161 benchmarks

In fitin the capability model (19 of 161)
Legal · 69 benchmarks · first place on 7
HLAB-Vals-1.1 - Bankruptcy Restructuring
1 / 36———
HLAB-Vals-1.1 - Employment Labor
1 / 36———
HLAB-Vals-1.1 - Environmental ESG
1 / 36———
HLAB-Vals-1.1 - International Trade Sanctions
1 / 36———
HLAB-Vals-1.1 - Trusts Estates Private Client
1 / 36———
Finance & Accounting · 36 benchmarks · first place on 1
Tax Agent Bench — Forms & Filings
1 / 77———
Finance Agent v2.1 - General Qualitative Analysis
3 / 36———
Finance Agent v2.1 - Comparables
6 / 36———
Tax Agent Bench — Citation-Adjusted Fact-Pattern All-Pass
6 / 77———
Tax Agent Bench — Rule & Source Lookup
8 / 77———
Code · 13 benchmarks
Terminal-Bench 4.0 — Media
7 / 55———
Terminal-Bench 4.0 — Software
11 / 55———
Terminal-Bench 4.0 — ML
11 / 55———
Terminal-Bench 4.0 — Operations
11 / 55———
APEX-SWEIn fit
12 / 65~2 mo~Jan 2028—
Indexes · 13 benchmarks
Vals Legal Index
4 / 31———
Vals Finance Index
9 / 36———
Vals Index v2.2
10 / 31———
Artificial Analysis Healthcare & Medical Index
10 / 23———
Artificial Analysis Finance & Accounting Index
11 / 25———
Agents · 8 benchmarks
OSWorld 2.1 full - reported binary completion, step limit unspecified
5 / 5——2026-11-09
AutomationBench-AA Finance
7 / 25———
AutomationBench-AA Support and Operations
7 / 25———
AutomationBench-AA Operations, HR, Marketing, Sales and Support
7 / 25———
τ³-Banking (Mercor run, bm25_grep retrieval)
8 / 37———
Professional Work · 8 benchmarks
APEX-1 (Mercor, mean score)In fit
10 / 68~5 mo~Sep 2028—
GDP.pdf (Artificial Analysis, Mean Criteria)x-high
15 / 157———
APEX-Agents 1.1In fit
15 / 63~3 mo~Dec 2026—
APEX-Agents 1.1 Pass@1
16 / 63———
APEX-1 (Mercor, Pass@1)
17 / 68———
General QA · 7 benchmarks
Humanity's Last Exam Diamond (Reasoning Partition)
19 / 95———
AA-Omniscience Health Accuracy
20 / 23———
AA-Omniscience Business Accuracy
23 / 25———
AA-Omniscience Law Accuracy
23 / 25———
Humanity's Last Exam Diamond
30 / 139———
Safety · 4 benchmarks · first place on 1
AA-Omniscience Health Non-Hallucination
1 / 23———
AA-Omniscience Business Non-Hallucination
3 / 25———
AA-Omniscience Law Non-Hallucination
3 / 25———
AA-Omniscience v1.1 Hallucination Ratex-high
4 / 68———
Science · 2 benchmarks
Terminal-Bench-Science 0.1In fit
16 / 70~1 mo~Mar 2027—
CritPt v2 (agentic, avg@5)In fit
21 / 69~2 mo~Feb 2027—
Health & Medicine · 1 benchmark
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)In fit
31 / 108~4 mo~Apr 2029—
Loading Atlas data…