DeepSeek V4 Flash 284B on AMD Strix Halo (gfx1151) — up to 32 tok/s decode & ~250 tok/s prefill via ROCmFPX, DSpark & ROCm 7.2
-
Updated
Aug 28, 2026 - C
#
localmaxxing
Here are 3 public repositories matching this topic...
Nanbeige4.3-33B on one RTX PRO 6000 Max-Q (325 W): 94.1 tok/s code, 105.3 with DFlash2. One-command recipe, every number a Verified LocalMaxxing run.
vllm local-llm llm-inference speculative-decoding sglang sm120 nvfp4 rtx-pro-6000 nanbeige localmaxxing
-
Updated
Nov 2, 2026 - Shell
Rent GPUs by the hour on Lium to top LocalMaxxing leaderboards with verified, high-speed inference for under $10 total.
reproducible amd autotest multi-gpu lium rocm testplatform cloud-gpu llm bittensor vllm local-llm llm-inference strix-halo gfx1151 decentralized-compute rtx-3060 dspark rocmfpx localmaxxing
-
Updated
Oct 13, 2026 - Shell
Add this topic to your repo
To associate your repository with the localmaxxing topic, visit your repo's landing page and select "manage topics."