SSyncDev

Muse Ember

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Muse Ember is Meta's smallest, cheapest Muse model: a 1M-context multimodal worker for tool calls, coding subtasks, routing and extraction at volume.

Key specifications

Specification score
81.8
Context window
1.05M
Max output
524.29K
Input / 1M tokens
From $0.10
Output / 1M tokens
From $0.50
License
Proprietary

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Indicative: prompts up to 200K tokens $0.10 in / $0.004 cached / $0.50 out per 1M; over 200K $0.30 / $0.012 / $1.50. Confirm on your own account.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input
  • Audio input
  • Video input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionAudio

What is Muse Ember?

Ten cents per million input tokens, fifty cents out, and a million-token window. Those three numbers explain why Meta built Ember: an agent that plans with a frontier model still makes hundreds of small calls underneath, and each of those calls should not cost frontier prices.

Meta describes its line-up in three tiers. Muse Wonder handles the hardest problems, Muse Spark covers coding and agentic workflows, and Ember is the repeatable, high-volume layer underneath. It is also the first model Meta says it trained using Muse Wonder Full. The suggested pairing is explicit: Wonder plans, Ember executes.

Spec sheet

Field Detail
Model ID muse-ember-1.0
Released 9 November 2026
Context / max output 1M / up to 512K tokens
Input modalities Text, image, audio, video
Output Text
Reasoning, tool calling Yes, both
Price ≤200K prompt (indicative) $0.10 in, $0.004 cached, $0.50 out per 1M
Price >200K prompt (indicative) $0.30 in, $0.012 cached, $1.50 out per 1M
Interfaces Meta Model API (Responses API), Muse Code

The long-prompt tier triples every rate, so parse a 600K-token contract once and cache it rather than resending it.

The jobs it was built for

Meta lists the work plainly: tool calls, coding subtasks such as reading files, running tests and applying targeted fixes, routing, extraction, and long-document parsing. Multimodal input stretches that to screenshots, charts, scanned documents, voice memos, meeting recordings and video, plus driving a computer from screenshots.

Publisher-reported results, read carefully

All of these come from Meta, run at maximum reasoning:

  • Terminal-Bench 4.0: 40.9%, against 39.2% for Claude Haiku 5.5 and 16.4% for GPT-6 Luna.
  • AutomationBench-AA: 37.8% versus 13.9% for Haiku 5.5.
  • Artificial Analysis Intelligence Index: 36, level with Haiku 5.5 and one point under Muse Spark 1.3.
  • FrontierCode 2.0-AA: 27.2%, behind Haiku 5.5 at 34.9%.

That last line is the useful one. On harder, open-ended coding, the cheaper model gives ground. Several of the other wins are on Meta-internal benchmarks, which nobody outside can rerun. Meta also puts its cost per Intelligence Index task at about $0.14, versus $2.39 for Spark 1.3.

Open questions before you commit

Meta describes native text, image, audio and video understanding; at least one early write-up mentioned only text, images and audio. Send a short video clip through your own key to settle it. Meta also never says outright that the weights are closed, though it is widely treated as a closed model, unlike the open Muse Glimmer 30B and Muse Spark Open. Assume no download.

Ember is served on the Meta Model API and Muse Code, and Meta says it is also offered through Google Cloud and Oracle Cloud.

A practical trial

Take a week of logged sub-calls from your existing agent, the boring ones: JSON extraction, "which tool next", a failing test fix. Replay them through Ember and through whatever you use today. Track three things: exact-match rate on structured output, retries per task, and dollars per completed task. If retries climb on coding fixes, keep the bigger model for those and route only extraction and routing down. Rates are indicative; confirm them on your own account. More from this lab on the Meta provider page.

How to evaluate Muse Ember

Compare the model on the work you need to complete.

Tasks and inputs

Muse Ember is listed for Multimodal, Text → Text, Coding & Code Generation, Computer Use. It accepts text, image, audio, video and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 1.05M context and 524.29K maximum output. The listed route has starting rates of $0.10 input and $0.50 output per million tokens. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.

Frequently asked questions

`muse-ember-1.0`, served through the Meta Model API (which supports the Responses API) and Muse Code.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

baidu

ERNIE 5.5 Flash Preview

ERNIE 5.5 Flash Preview is Baidu's first post-trained 280B-A14B text model, with a 1M window, aimed at game dev, 3D scripting and knowledge work.

1M context

iflytek

Spark X2.6

Spark X2.6 is iFLYTEK's 293B mixture-of-experts language model family with integrated device and cloud variants, unveiled in October 2026.

— context

Meta

Muse Spark Open

Muse Spark Open is Meta's open-weight 816B mixture-of-experts model (42B active) with 1M context, built for coding and agents you can self-host.

1.05M context

ByteDance Seed

Seed 2.1 Lite

Seed 2.1 Lite is a lightweight Doubao Seed 2.1 model listed on Volcengine Ark with a 1M context, 256K output, audio and image input and tool calls.

1.02M context