I use the R1 variant on my local only machine. It’s great for logic and analysis. Depending upon speed vs accuracy, I use the 8B, or 32B variants.
Deepseek is definitely worse than the best ChatGPT and Anthropic models. It’s especially evident in coding tasks, but it is also worse with hallucinations and reasoning generally, in my experience. But for basic language tasks or pointing me to research on a topic, it’s pretty good.
I just don’t see it unseating the big american firms without an additional push.
I think V4 is comparable to Claude when it comes to coding, but it hasn’t been shipped to the web version yet.
Oh nice. I may have to try it out
A lot of that has little to do with model capability and comes down to coding harnesses not meeting the expectations of the model. Here’s a great discussion regarding that https://xcancel.com/MrAhmadAwais/status/2050956678502420612
DeepSeek team is aware of the tooling gap and now they’re working on their own harness to close it https://deepseekv4pro.com/news/deepseek-code-harness-team-claude-code-rival-report
Whoa cool. Thanks for the info
“According to Chinese newspaper, US tech firms can’t afford the dollar costs and prefer paying in IP”(intellectual property)
edit: FYI source of Yahoo for this article is the South China Morning Post.
that’s good to know, i much rather get my news from a reliable source than from a western propaganda one.
