Muse Ember vs Muse Spark 1.2
Muse Ember and Muse Spark 1.2 score 76.6% vs 73.8% averaged across five Vision Evals tasks (#22 and #26 of 73 models), so Muse Ember leads overall.
- Muse Ember wins all 5 tasks, with its widest lead on object detection (63.8% vs 59.0%).
- Muse Ember is about 1/9 of the cost: $0.0009 vs $0.0083 overall average cost per sample.
- Muse Ember is faster: 6.1s vs 11.5s per sample.
Roboflow Vision Evals. Full breakdown below.
Compare Muse Ember vs Muse Spark 1.2 live
Run the same image across every model that supports a task and compare their outputs side-by-side.
These models don't share enough common tasks for a side-by-side demo. See the comparison table below for their capabilities.
Muse Ember vs Muse Spark 1.2 on Vision Evals
Roboflow Vision Evals. Full breakdown below.
Muse Ember vs Muse Spark 1.2 Comparison Table
Evals updated November 19, 2026Pricing updated November 21, 2026
| Property | Muse Ember | Muse Spark 1.2 |
|---|---|---|
| Organization | Meta | Meta |
| Category | closed | closed |
| Modality | multimodal | multimodal |
| Release Date | Nov 2026 | Aug 2026 |
| Context Window | 1.0M | 1.0M |
| Parameters | Unknown | Unknown |
| License | Proprietary | Proprietary |
| Commercial use | commercial use per vendor terms | commercial use per vendor terms |
| Pricing per 1M tokens | ||
| Input $/1M | $0.100 | $1.25 |
| Output $/1M | $0.500 | $4.25 |
| Vision Tasks | ||
| Captioning | Supported | Demo |
| Classification | Supported | Demo |
| Object Detection | Supported | Demo |
| OCR | Supported | Demo |
| Visual Question Answering | Supported | Demo |
| Model Features | ||
Both support 3 more features | ||
Vision Evalsground-truth scores across 5 vision tasks, pooled at low effort | ||
| Overall | 76.6% | 73.8% |
| Avg cost / sample | $0.0009 | $0.0083 |
| Avg speed / sample | 6.08s | 11.51s |
| By task | ||
| Object Detection (low) | 63.8% ±0.5, Mean of 3 runs, range 63.3 to 64.3 | 59.0% ±1.0, Mean of 3 runs, range 58.1 to 60.2 |
| Object Detection (high) | 64.0% ±0.5, Mean of 3 runs, range 63.5 to 64.5 | 60.5% ±0.3, Mean of 3 runs, range 60.2 to 60.7 |
| Counting (low) | 79.7% ±1.4, Mean of 3 runs, range 78.4 to 81.1 | 76.6% ±2.7, Mean of 3 runs, range 74.3 to 79.7 |
| Counting (high) | 79.3% ±0.7, Mean of 3 runs, range 78.4 to 79.7 | 75.2% ±2.0, Mean of 3 runs, range 73.0 to 77.0 |
| Identification (low) | 91.7% ±1.6, Mean of 3 runs, range 90.6 to 93.8 | 89.6% ±1.6, Mean of 3 runs, range 87.5 to 90.6 |
| Identification (high) | 90.6% ±3.1, Mean of 3 runs, range 87.5 to 93.8 | 87.5% ±0.0, Mean of 3 runs, range 87.5 to 87.5 |
| OCR (low) | 69.5% | 68.7% |
| by category |
|
|
| OCR (high) | 70.5% | 68.6% |
| by category |
|
|
| Reasoning (low) | 78.2% ±1.3, Mean of 3 runs, range 76.8 to 79.5 | 75.1% ±0.3, Mean of 3 runs, range 74.8 to 75.5 |
| Reasoning (high) | 81.9% ±1.7, Mean of 3 runs, range 80.1 to 83.4 | 75.7% ±0.3, Mean of 3 runs, range 75.5 to 76.2 |
Muse Ember vs Muse Spark 1.2: Overview
Muse Ember is a proprietary multimodal reasoning model from Meta, released on November 9, 2026. It accepts text, images, video, audio, and PDF documents and returns text. The Meta Model API serves it as muse-ember-1.0, with a 1,048,576-token context window and up to 524,288 output tokens. Reasoning effort is adjustable from minimal to max.
Meta positions Ember for repeated tasks within larger agent workflows: tool calls, routing, coding subtasks, and processing collections of images and documents. It can caption or sort photos and extract fields from records while a larger Muse model handles planning. Meta recommends Spark or Wonder for harder single-shot coding and knowledge-heavy problems. Ember is available as a hosted model; its weights are not publicly released.
Muse Spark 1.2 is a proprietary multimodal reasoning model from Meta Superintelligence Labs, released as a coding-focused update to Muse Spark 1.1. It accepts text, images, video, audio, and PDF documents and returns text, with a context window of roughly one million tokens that allows whole repositories, long documents, and extended agent trajectories to be held in a single request. The model thinks before answering, and the amount of reasoning effort it spends is configurable per request. Alongside its visual and document understanding, it supports structured output and parallel function calling, and it is designed to operate either as a planning agent that delegates work or as a subagent executing tasks in parallel.
Training for version 1.2 scaled up compute on coding tasks and widened the diversity of training environments, concentrating on long-horizon work such as whole-repository generation, large end-to-end projects, and automated research. Part of the training data was self-generated, with Muse Spark 1.1 producing coding environments and instruction-following templates and grading candidate solutions against them. The model was co-trained with the Muse Code terminal agent, incorporating rejection-sampled harness trajectories and that toolset. Multimodal use cases documented for the family include visual-to-code generation and detailed image and video captioning.