ling-30-flash
candidate
citable URL: https://halobench.com/records/ling-30-flash/ — this address never moves; the anchor /records/#ling-30-flash keeps resolving
Ling-3.0-flash @ Q4_K_M corrected GGUF · kind model · engine igpu · tier candidate
runs on aihydra · config cfg-0115
⌁ current state candidate · ⌁ days in production 0
Lifecycle — append-only
2026-08-18
candidate · run-0439
Corrected bailingmoe3 GGUF screened and benched after llama.cpp PR #26608 landed. Text-only Ling family; multimodal capability lives in separate Ring/Ming lines rather than this checkpoint. First publishable full bench is ROCm plain decode only, no speculative/MTP path exercised.