Hacker News Reader

Comment by incone123 | Hacker News Reader

Surely investors and tech leaders, as distinct from speculators, should be testing the product and doing due diligence on the company. I'm just a layman and I liked how a small version of deepseek ran faster and with less power draw on my PC than other small version models did. Output, very informally tested, was good. I did not think the 'reasoning' text added anything useful but I'm also not qualified to say how it contributed to the final output.

View on HN

> I'm just a layman and I liked how a small version of deepseek ran faster and with less power draw on my PC than other small version models did.

This doesn't sound accurate. There are no "small versions" of deepseek r1. They released some distilled fine-tunes of their big model. They took qwen2.5 models as a starting point, and fine-tuned them with ~800k generations made with deepseek r1. But they did not change any of the model's architecture, so the "base" model released by qwen and the fine-tuned one released by deepseek should run identically on your hardware.

Emphasis on layman and I'm not at my PC but when deepseek came out there was something I could pull in ollama that could run on modest hardware.

Yes, that's ollama's fault. They renamed the fine-tuned qwen model to deepseek, and a lot of people got confused. But behind the naming blunder, the model you got was qwen2.5-finetuned-r1, which was identical in architecture with qwen2.5 (so it should have ran at the exact same speed)

Reply on news.ycombinator.com