Reddit user reports strong Qwen 3.8 Flash Next performance with Basalt
AI summary
A Reddit user reports running the NVFP4 version of Qwen 3.8 Flash Next through Basalt on CachyOS with 64GB of RAM and an RTX 5090. The user reports 200 tokens per second decoding, about 9,000 tokens per second prefill, and a 256K context, while saying that Windows 11 memory overhead reduced performance.
In-depth insight
The post provides a first-hand local-inference configuration and performance data for a quantized Qwen model, offering practical evidence for users running large models on high-end consumer hardware.