gittensor-model-hub/Qwen3.8-27B-DSpark-NVFP4

Fix documented --max-mamba-cache-size (4 -> 8): 4 left no headroom

Update card for v3 drafter: agentic +14.2%, held-out +4.5%

v3 drafter: config.json

v3 drafter: retrained on serving-format corpus (+14.2% agentic acceptance)

NVFP4 mixed-precision DSpark drafter: 147.87 tok/s, 2.792 accept, 1.41 GB

initial commit