Mesh Applied Optimizes Qwen3.8 for 255 tokens per second on RTX 5090
October 13, 2026
Mesh Applied uses quantization and kernel specialization to achieve 255 tok/s on a single RTX 5090 with Qwen3.8-Flash-Next-NVFP4. Their Goethe framework also enables DeepSeek-V4.1-Flash-Ternary to reach 276 tok/sec aggregate on 4x RTX6000 hardware.
HOW THIS AFFECTS YOU
●
builderYou can achieve high-throughput local inference for coding tasks using specialized quantization kernels.
●
researcherYou can explore the performance gains of kernel specialization and NVFP4 quantization on consumer-grade hardware.