Muse Spark Open vs Muse Spark 1.3 and the leading open-weight models

Muse Spark Open is Meta's open-weight release of Muse Spark, tuned for agentic coding and long-running agent work. It is a mixture-of-experts model with 816B total and ~42B active parameters. It takes text, image, audio and video input, with a 1M-token context. The weights are on Hugging Face: about 1.6 TB in BF16, or about 530 GB in NVFP4, which Meta says serves on four Blackwell Ultra GPUs or eight H200s.

Same API price as Muse Spark 1.3: $1.25 / $4.25 per 1M tokens, $0.15 cached input.

How it compares (Artificial Analysis Intelligence Index):

• Muse Spark 1.3 48.1 · DeepSeek V4.1 Flash 1017 45.1 · GLM 5.3 44.8 · Kimi K3 43.6 · Qwen3.8 2.4T A95B 39.9 · Muse Spark Open 38.9
• On the Index it sits about 9 points below the closed Spark 1.3 at the same price. The value of Spark Open is the weights, not the hosted API.
• Output speed is ~142 tokens/s, against ~133 for Spark 1.3 and ~42 for Kimi K3.

From Meta's own model card: on Terminal-Bench 2.1 it scores 76.4, behind Kimi K3 (85.0), GLM 5.3 (83.9) and Qwen3.8 2.4T (82.0). Meta's comparison runs Muse Spark Open at xhigh effort and Kimi K3 at max.

Licensing: the weights are under the Muse Model License, so check its terms before you build a product on them.

Build with Muse Spark Open (Meta) · Model card