SF-Foundation/Koa-ultra-nvfp4

Koa-ultra NVFP4

Built with Muse Spark Open

This checkpoint merges SF-Foundation/Koa-ultra-adapter into the published Muse Spark Open 816B-A42B NVFP4 checkpoint. It contains the full model weights; loading a separate LoRA adapter is unnecessary. See merge_provenance.json for pinned source revisions, merge details, and numerical checks.

The Fireworks rank-8 shared-outer expert factors are merged at alpha/r = 4 using FP32 accumulation. All 62 exported sampler routing-bias states are restored. Quantized expert matrices are dequantized, merged, and re-encoded using NVIDIA ModelOpt NVFP4 with new FP8 block scales and FP32 global scales. The original BF16 exclusions, multimodal weights, tokenizer, and input activation scales are retained. No new activation calibration was performed.

This process introduces re-quantization error. Structural and numerical merge checks passed; end-to-end generation and task quality have not been evaluated. Use a runtime supporting the Muse Spark architecture and ModelOpt NVFP4 format.

The Muse Model License and Usage Policy are included in LICENSE.md and USE_POLICY.md. This derivative remains subject to them.