日月小楚|12月 03, 2025 13:19
Stepped on a big pit today! How did I not think that the same model could perform so differently?
Yesterday, after the official release of deepseek 3.2, its capabilities were on par with GPT-5. So, I handed over part of my daily tasks to it. It handled over 30 analysis reports, but 8 of them had serious issues.
This result was really unexpected because it was way off from what I anticipated. So today, I connected to the official API and ran it again.
This time, the results were excellent. It truly performed at GPT-5's level, completely different from the previous run.
After thinking it through, the problem lies with the first run using OpenRouter's API. As you all know, there are so many large models now, and if you have to register, top up, and configure the call codes for each one, it’s super tedious. OpenRouter integrates almost all the major models, which is especially convenient when testing new ones.
The model supplier for deepseek V3.2 provides FP4 and FP8. FP4 and FP8 refer to floating-point data precision formats. Simply put, they are techniques to "compress" models (i.e., quantization) to make them run faster and use less memory, while trying not to compromise intelligence.
Although for most models, FP4 and FP8 almost don’t affect intelligence, the DSA architecture used by deepseek might be significantly impacted by this.
Man, the world of AI is still way too complicated.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink