stealthmodels.

MODEL BENCHMARKS

Muse Spark Open: Newton's Mirror

Meta's open-weight coding and agent model combines 816B total parameters, 42B active per token and native image, audio and video understanding.

A good fit for

Coding agents, repository analysis and multimodal workflows, with hosted APIs or Unsloth GGUFs for local deployment.

Context window
1,048,576tokens[7] [8]
Active parameters
42Bparameters[7]
License
Muse Model License[4]
Model factsArchitecture, context & capabilities
First available
2026-10-15[2]
Developer
Meta[2]
Origin
United States[3]
Family
Muse Spark[1]
Context window
1,048,576 tokens[7] [8]
Maximum output
1,048,576 tokens[8]
Architecture
Sparse mixture-of-experts Transformer with perception encoders; 62 decoder layers, 256 experts with 8 selected per token, grouped-query attention and sliding-window/full-attention layers.[5] [6]
Total parameters
816B[7]
Active parameters per token
42B[7]
Input
text, image, audio, video[7]
Output
text[7]
Normal speed
361.08 tokens/sec avgMin 290.98 · Max 424.52 tokens/secOur measurement
API access & pricingVercel AI Gateway · $1.25 input / $4.25 output · per 1M tokens

Vercel AI Gateway / OpenRouter; input 1.25 USD/1M; cached read 0.15/1M; output 4.25/1M

[8] [10] [13]
  • Vercel AI GatewayObserved 2026-10-15
    Input $1.25Output $4.25Cached read $0.15

    per 1M tokens

    meta/muse-spark-open[8]
  • OpenRouterObserved 2026-10-15
    Input $1.25Output $4.25Cached read $0.15

    per 1M tokens

    meta/muse-spark-open[10]
Run it locallyWeights, memory & deployment

High-memory local or multi-GPU server; Unsloth quantization and CPU/GPU offload supported

[11]
  • UD-IQ1_M208.93 GB storage · UD-IQ1_M[12]
  • UD-IQ1_S175.64 GB storage · UD-IQ1_S[12]
  • UD-IQ2_S275.5 GB storage · UD-IQ2_S[12]
  • UD-IQ2_XXS225.57 GB storage · UD-IQ2_XXS[12]
  • UD-IQ3_XXS340.35 GB storage · UD-IQ3_XXS[12]
  • UD-Q2_K_XL311.23 GB storage · UD-Q2_K_XL[12]
  • UD-Q3_K_XL401.48 GB storage · UD-Q3_K_XL[12]
  • UD-Q4_K_XL505.5 GB storage · UD-Q4_K_XL[12]
Measurements & sources13 primary references

StealthMark score

StealthMark measures fluid and visual intelligence of AI models, resisting any pre-existing benchmark focused training.

General and Visual

General Intelligence

The ability to reason, solve problems and understand how things relate and behave.

Visual Intelligence

The ability to imagine, create and accurately reconstruct complex images in code form.

Judging

Scores reflect how well each result meets the benchmark requirements.

Overall score

Overall combines weighted task scores on a scale from 0 to 100. Estimated means the total includes an estimate for one missing scene.

Scoring categories

Task & likeness

Subjects, actions and likeness.

Anatomy & construction

Coherent bodies, joints and machinery.

Space & placement

Perspective, scale and contact.

Light & reflections

Lighting, shadows and reflections.

Purposeful detail

Clear, purposeful small features.

Artistry & character

Composition, expression and drawing skill, including complexity and its execution.

Materials & effects

Surfaces, texture and movement.

Visual integrity

Clean shapes, layers and edges.

BENCHMARK RESULTS

Compare results

This link opens the same results in the same order.

StealthMark benchmark results

Reset benchmark preferences?

Return to the default view, clear comparison selections and restore the advertisement.