［ninfer v100X2］v100 achieves new glories! qwen3.8 27b nvfp4 2000t/s prefill, 115t/s decode
［Ninfer V100X2］v100 achieves glory again! qwen3.8 27b nvfp4 peak 236t/s decoding, 2000t/s prefill
The strongest 4-card V100 inference engine in history! Number one in the world, number two in China!
God-level project Strata! Two V100s running Qwen3.8 Flash IQ3_S, tested at 60-100 tokens/s
Tesla V100 PCIE version and P100 PCIE version modified active cooling solution, with 4090 mod and water cooling available, and can ensure single-slot multi-card setup.
V100 unlocked NVFP4, GLM5.3F insane output 145 toks
Local model Qwen3.8 (choose 27B or flash-next-strata)
[ninfer-V100X2] V100 achieves glory yet again, real-world usage non-benchmark scenario 183t/s decoding, 2500t/s prefill
V100 Duo! Two V100s running NInfer NVFP4, 27B model breaks 110 tokens/s?
The sky is falling! V100 + 3080 running Strata on an X99 motherboard hit an unbelievable 197 tokens/s! Includes a Modpack with one-click download, start your own AI life
The Best Cost-Effective AI Local Deployment in 2025! Dual V100! 32G VRAM, High Performance at Low Price, Crushing 2080Ti 22G!
Four 170HX + DDR3, fully loaded DeepSeek V4.1 Flash 1017 running at 50+ tokens/s for a single stream!
Qwen3.8 27B, 256K context, running at 60 tokens/s on a single 16G V100. Come and try it out!
Strata model selection: run out of memory and go to SSD? First finish reading this downgrade table
[Ninfer V100X2] Peak 68t/s! Dual V100 16g running qwen3.8 27b q4 quantization 180k context, achieving 68t/s decoding?
【刚哥玩生态】双V100 Nvlink模型服务器起飞了
strata memory 32G + 8G VRAM > 39G is enough to run an IQ3-level 85G file without reading the hard drive at an average speed of 66t/s - Gyro-M
No words needed, just look at the picture. Dual V100 64G running Qwen 3.8 Flash NEXT. 256K context, peak 190 tokens/s.
[Most Comprehensive Test] Deploying Qwen3.8 27B nvfp4 on V100 with vLLM + dFlash2: Decode speed improved by 3x, Prefill speed improved by 15x
2080Ti 22g + 48g RAM running Qwen3.8-Flash-Next (IQ3_XXS): stable 45+ tok/s, amazing results
