hanxiao/DeepSeek-V4.1-Flash-1017-abliterated-4x-RTX-PRO-6000: Serving setup and benchmarks

An engineer details the configuration and deployment of an abliterated DeepSeek-V4.1-Flash model served on a Google Cloud Compute Engine VM with four RTX PRO 6000 GPUs. Utilizing vLLM-Moet, DSpark speculative decoding, and FLy verification, the setup achieves 463 tokens per second on single-stream agent decode and 1,082 tokens per second across eight streams. Open article →