Qwen3.8-Flash-Next hits 184 tok/s on one RTX 6000 Pro

Mesh Applied released an NVFP4-quantized Qwen3.8-27B-NInfer model, claiming 184 tok/s on a single RTX 6000 Pro. The team argues open models already suffice for about 98% of real AI tasks, so the bottleneck is running them often and fast, and it optimizes quantization and kernels around time per task.

• Qwengram-0.8B: n-gram memory transfer cuts perplexity 5.05%2026-09-25
• Microsoft releases 4B vision-language model Mage-VL-Cowork2026-10-31
• DFlash2 drafter lifts Qwen3.8-Flash-Next speed by 17%2026-10-31
• Developer builds Halloween haunted house with NVIDIA Nemotron models2026-10-30
• Community releases Gemma-4-31B-Humanlike-Chat fine-tune2026-10-30