Releases

Muse Spark Open

MiniRouter and Muse Spark Open by Meta, separated by a silver fold of light on black.

Meta releases open weights for coding and agent workflows, with 816B total parameters and 42B active per token.

Muse Spark Open is available on MiniRouter as . The model page has current pricing, provider routes and API setup.

View Muse Spark Open

Spark, with the weights in your hands

Build coding agents, work across long inputs, or adapt the model to your own tasks.

Architecture
816B total · 42B active
Sparse mixture of experts; active parameters are per token.
Context
1,048,576 tokens
Input and output share the context window.
Weights
Muse Model License
Download the weights to adapt and host the model.
Reasoning
Low through xhigh
Choose low, medium, high or xhigh effort.

A large model. A small active slice.

42Bactive per token
About 5.1% of the model activates for each token. A sparse mixture of experts selects which parameters do the work.

Architecture and hosting details

Strong tool use. A mixed coding picture.

Spark Open leads this group on MCP-Atlas. On terminal tasks, Qwen, GLM and Kimi score higher.

MCP-Atlas

Complete tasks by finding and using the right tools.

  • Spark Open89.2%
  • Qwen 3.884.5%
  • GLM 5.384.2%
  • Kimi K382.3%

0–100% · higher is better

Scale AI's official agent harness. Tasks pass at 75% claim coverage or higher.

Meta's published results · 15 October 2026. Spark Open and Qwen use xhigh; GLM and Kimi use max. Results · Methods
View chart data
MCP-Atlas comparison data
ModelEffortScore
Muse Spark Openxhigh89.2%
Qwen3.8-2.4T-A95Bxhigh84.5%
GLM 5.3max84.2%
Kimi K3max82.3%

Terminal-Bench 2.1

Solve coding and system tasks in a terminal.

  • Spark Open76.4%
  • Qwen 3.882.0%
  • GLM 5.383.9%
  • Kimi K385.0%

0–100% · higher is better

Artificial Analysis evaluation with the Terminus 2 agent in E2B, across 89 tasks.

Meta's published results · 15 October 2026. Spark Open and Qwen use xhigh; GLM and Kimi use max. Results · Methods
View chart data
Terminal-Bench 2.1 comparison data
ModelEffortScore
Muse Spark Openxhigh76.4%
Qwen3.8-2.4T-A95Bxhigh82.0%
GLM 5.3max83.9%
Kimi K3max85.0%

Give the answer room to land

Choose low, medium, high or xhigh effort. Keep enough output tokens for both reasoning and the final answer.

Your prompt

One output budget

ReasoningAnswer
Start at medium for everyday work. Try higher effort for a difficult bug or a plan with several dependencies.

Put it to work

Try a concrete coding task, then check the result against your own examples.

Muse Spark Open · cURL
curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "meta/muse-spark-open",
    "reasoning": { "effort": "high" },
    "max_tokens": 8192,
    "messages": [{
      "role": "user",
      "content": "Write a CSV parser in Python that handles quoted commas and escaped quotes. Include examples for each edge case."
    }]
  }'

Current prices and client setup →

Keep exploring

Compare published results in the model rankings and coding benchmarks, or read more model releases.

For MiniRouter product updates, visit the changelog.

Sources

  1. Meta Muse Spark Open announcement

    Published Oct 15, 2026

    Read at Meta ↗
  2. Muse Spark Open model card

    Published Oct 15, 2026

    Read at Meta ↗
  3. Muse Spark Open evaluation methods

    Published Oct 15, 2026

    Read at Meta ↗
  4. Muse Spark Open architecture and hosting

    Published Oct 15, 2026

    Read at Meta ↗
  5. Vercel Muse Spark Open catalog

    Published Oct 15, 2026

    Read at Vercel ↗

Editorial reference revised 2026-10-16. Catalog status comes from the generated routing snapshot.

Every model, one key.