AnalysisAI ModelsJuly 6, 2026
REAP-pruned DeepSeek-V4-Flash runs on single Spark GX10

User ran a 162B REAP-pruned NVFP4 DeepSeek-V4-Flash model on a single Ascent GX10 Spark by patching a custom vLLM docker image. The setup maintained consistent performance at long context.