From SiliconANGLE · Duncan Riley · · 1m 10s
Meta Releases 816B-Parameter Muse Spark Open
Meta launches open-source AI model for coding and agent work with 816 billion parameters.
In brief
Meta launches open-source AI model for coding and agent work with 816 billion parameters. Meta released an open-weight AI model with 816 billion parameters for coding and agent tasks, outperforming most Chinese models in tool use but lagging in coding benchmarks. Originally reported by SiliconANGLE.
Meta releases Muse Spark Open
Meta Platforms Inc. released open weights for its Muse Spark AI model designed for coding and long-running agent work.
Model specs
The model has 816 billion parameters with only 42 billion switching on per token for coding and agent tasks.
Context window
Prompts can mix text, images, audio, or video with a context window up to 1 million tokens.
Safety features
Meta added safety post-training to hold up without Meta’s own services guardrails.
Benchmark results
Muse Spark Open scored 61.8 on JobBench agent test, more than twice Inkling’s 31.1.
Chinese models stronger in coding
Kimi K3 led Terminal-Bench 2.1 coding test at 85, while GLM 5.3 and Qwen3.8 scored above Muse Spark Open’s 76.4.
Tool use performance
Muse Spark Open scored 89.2 on MCP-Atlas tool use, beating all rivals including Chinese models.
The takeaway
Meta released an open-weight AI model with 816 billion parameters for coding and agent tasks, outperforming most Chinese models in tool use but lagging in coding benchmarks.
