Meta's Muse Spark Open: 816B total params, 42B active, 1M-token context. The license reads like MIT, then adds three conditions:
1. Products must show "Built with Muse Spark Open"
2. Derivatives include models trained or distilled on its outputs
3. API businesses past $20M revenue in 12 months need a Meta agreement
All 816B weights still sit in memory (~700 GB VRAM at NVFP4).
Meta released the weights of Muse Spark Open on Oct 15: 816B parameters, 42B active, 1M-token context, built for coding agents.
Easy to miss: "active" parameters set compute per token, but memory must hold all 816B.
And "open-weight" isn't open source: the license has limited commercial terms. Read it first. Benchmarks are Meta's own.
NVIDIA released the full Run:ai GPU orchestration platform as open source on Oct 13.
Worth noting: KAI Scheduler can be adopted alone, and fractional GPUs now rely on CUDA 13.4 native support rather than the old proprietary code. The NGC Community build has no NVIDIA support.