Muse Spark Open EXL3 K4 + Core K6

Built with Muse Spark Open.

Quantized from the original BF16 checkpoint meta-models/Muse-Spark-Open-816B-A42B at immutable revision 0fb02478b02565b2ce7fc57e6051c91172774ac9. The original Muse Model License and Usage Policy are retained as LICENSE.md and USE_POLICY.md.

Total tensor payload is 427,855,788,244 bytes (398.471754 GiB), 63 shards and 193,459 tensors. Safetensors headers and metadata add a small overhead. There is no MTP or shared expert.

Expert source gate_up_proj is split on output axis 1 after selecting expert axis 0: gate rows 0..4095, up rows 4096..8191. Published modules use model.language_model.layers.N.mlp.experts.E.{gate_proj,up_proj,down_proj}. The shared per-layer experts.post_expert_norm.weight remains byte-identical. Its norm is applied separately to each expert output before routing weights.

All whole K6 operands use sharded_side=1. lm_head has 202,048 logical outputs and 202,112 stored outputs; the final 64 output scales/rows are zero. Runtime trims these outputs before the original final softcap. Every stored extent, bit width and buffer specification is in quantize_config.json:tensor_storage; the safetensors headers independently encode all packed shapes. MCG multiplier is 0xCBAC1FED, automatic output scales, seed 787 and sigma 0.025. The checkpoint requires native EXL3 MCG K4/K6 support, plus the preserved native Muse components.

Calibration uses the pinned public agentic panel with the checkpoint's own chat template/tokenizer. Exact prompt counts, token counts, rendering settings, source revision, producer code SHA-256 values and numerical recipe are in quantization-provenance.json and quantize_config.json:calibration. A causal interleaved pass propagates serialized/reconstructed predecessor weights before calibrating each successor; routed down inputs follow the reconstructed gate/up nonlinear activation. The original forward preserves Muse attention norms/gating, routing, residual gates and per-expert norms.

Quantization used only raptor's GPU 0 (RTX PRO 6000 Blackwell). Independent validation compared all retained tensors byte-for-byte against source and verified exported shapes, scales, multiplier, tensor inventory and payload sizes. Complete per-module numerical evidence is bundled under quantization-errors/. This quantization job has not qualified full-model generation, long-context quality or serving performance. Vision/audio components are preserved, and were not calibrated or quantized by this job.