Pyyan / Compare / Qwen3.8-27B vs Muse Spark Open

Qwen3.8-27B vs Muse Spark Open

2 of 5

Open-Weight Models · verified 20 Oct 2026

Qwen3.8-27BAlibabacurrent
Muse Spark OpenMetacurrent
3 slots left
SpecificationQwen3.8-27BMuse Spark Open
Summary27B dense under Apache 2.0, and nearly five million downloads.The largest American open weights, under a licence with a revenue trigger.
Context262K native, ~1M extended1M
Input ($/Mtok)Weights onlyWeights only
Output ($/Mtok)Weights onlyWeights only
LicenceApache 2.0Muse Model License
Max output32KNot published
Parameters27B dense816B total, 42B active
Released5 Aug 202615 Oct 2026
Model IDQwen/Qwen3.8-27Bmeta-models/Muse-Spark-Open-816B-A42B
CategoryOpen-Weight ModelsOpen-Weight Models
Official—Meta ↗

Highlighted rows are where these differ.

Qwen3.8-27B

  • 27B dense rather than a mixture of experts, so it behaves predictably on one machine
  • Apache 2.0, weights only, no hosted price
  • One of the most downloaded models on Hugging Face, with quantised builds everywhere

Best for running locally.

Full spec sheet →

Muse Spark Open

  • Text, image, video and audio in, text out, with a 1M token context
  • Muse Model License, not open source: attribution required, and a separate agreement above $20m of model-service revenue
  • About 1.5TB in BF16, or two Mac Studios at 4-bit

Best for multimodal work you host yourself.

Full spec sheet →