Community
Basalt 0.2.0 adds Qwen 3.8 Flash Next NVFP4 support with faster local inference on Blackwell GPUs
The update reports about 300 tokens/s for prose and 600 tokens/s for structured output on an RTX 5090 with a 5060 Ti, and adds a higher-quality NVFP4-DQ8 build with KLD 0.0561 versus 0.0624 for the DQ6_K default. Speed figures come from the author's own rig.