NVIDIA’s FP4 QAD Quants Reach the DGX Spark

NVIDIA has released a new FP4 quantization of GLM 5.3 Flash made with quantization-aware distillation (QAD), @MiaAI_lab reported on October 16. It fits on two DGX Sparks and on two RTX PRO 6000s, and per NVIDIA it keeps almost all of the model’s BF16 quality.

Owners of a single DGX Spark got a similar boost. @MiaAI_lab says Qwen3.8 Flash Next can now run on TensorFold with NVIDIA’s new FP4 QAD quant, with 1.1M total FP8 KV cache, 256k context and five concurrent users. She measured 46 to 55 tokens per second on prose in a single stream and 132 tokens per second across five concurrent streams.

A day later she reported further improvements: roughly a 14% uplift in prefill across the board for Qwen3.8 Flash Next FP4 QAD, with no change to decode speed for now.