@MiaAI_lab Finds Flare Short of a Posted Benchmark

No EXL3, for now.

This morning @MiaAI_lab reported NVIDIA’s release of Flare, a local inference server for DGX Spark and RTX Blackwell GPUs. A few hours later she had a first gripe: “Seems like it doesn’t support EXL3.”

Testing someone else’s numbers.

A user posted results for GLM-5.3-Flash on 2× DGX Spark with Flare: 117.2 tok/s for single-stream prose decode, 167.4 tok/s aggregate with 2 streams and 215.8 tok/s with 4, plus 2,327.9 tok/s for 32K prefill.

“You’re getting 100+ on prose on 2 sparks?” @MiaAI_lab asked. After running it herself, she replied: “I tested it myself and it’s nowhere near those speeds.”