Currently, both are running on two DGX Spark 2 nodes. The current connection is not direct but connected to a switch on MIKRO CRS 504.
QWEN 3.8FN is overwhelmingly faster.
However, the problem is that QWEN 3.8FN gets stuck in a strange loop every 30 minutes. It's a bit of an incomprehensible situation. It keeps printing "!!!!" characters and generating meaningless tokens. Turning off Thinking seems to prevent it from happening, but I need to confirm it further.
GLM 5.3 F was also run under the same conditions, and its speed doesn't seem bad compared to the toks measurement.
As a quality test, when I gave a one-line prompt "Make Tetris," the result was surprisingly better with QWEN 3.8 FL.
Both had bugs, but for QWEN, there was one where the NEXT block wouldn't update, and
for GLM, there were three: the NEXT block wouldn't update, a single line would remain when deleting multi blocks, and the hold block wouldn't display. Of course, you can't judge based on this alone, but I was a little disappointed because I thought GLM would be better.
And for the next one-line prompt "Make Galaga," GLM seemed to understand the Galaga game better. ㅎㅎ I was surprised to see it even adding sound effects.
It's only been about a day since they were released, so I need to do more testing, but one thing is for sure: both are undeniably pushing the frontiers.
It's quite amazing. ㅎㅎㅎ
Just take it as a reference for fun~
| QWEN 3.8 FLASH NEXT NVFP4 | GLM 5.3 FLASH NVFP4 |
|---|
Single stream | 47.6 tok/s | 21.9~33.3 tok/s |
TTFT | 0.16 seconds | 0.43 seconds |
Concurrent 4 (aggregate) | 119.1 tok/s | 50.6 tok/s |
Streams per @4 | 29.8 | 12.6 |
QWEN3.8 FLASH NEXT
GLM 5.3 FLASH
