REVIEWS / AI MODELS / OWNER INSIGHTS

🦉 WE READ 325 OWNER COMMENTS

DeepSeek Chat: what owners actually say

Owners praise DeepSeek's open-source power and cost, but repeatedly hit a wall on Chinese-state censorship and reliability issues.

LEMMY · 235 YOUTUBE · 52 HACKERNEWS · 29 STACKEXCHANGE · 8 PRODUCTHUNT · 1

What owners complain about

  • Censorship on China-sensitive topics COMMON

    Numerous users report the model refuses or evades questions about Tiananmen Square, Taiwan, and Uyghur abuses. It states topics are 'sensitive in China,' says 'there is no country called Taiwan,' or deflects with 'let's talk about something else.' R1 is noted as particularly evasive compared to V3.

  • Propaganda concerns SOME

    Users warn that R1 and V3 'deny the tiananmen massacre, the abuses on uyghurs, and defend the CCP.' Multiple commenters express alarm that LLMs replacing search could normalize state-driven narratives for a generation.

  • Server capacity issues FEW

    Users report frequent 'Server busy' errors when trying to access the service.

  • Verbosity FEW

    Users mock the model for over-producing content, e.g. 'about to give a 20 paragraph essay on how grass is green.'

  • Cannot run locally without major hardware FEW

    One user pushing back on 'open source = run at home' points out you need 'thousands of almost top of the line graphics cards' and 'millions of dollars to pay for electricity' to train or retain the model from scratch.

What owners love

  • Genuinely open source under MIT

    Users repeatedly highlight that code, weights, and model are released under MIT license, enabling local deployment, fine-tuning, and modification — praised as 'a great boon.'

  • Performance rivals or beats Western models

    Multiple users say they have 'completely switched to using DeepSeek for everything' and find it 'better than Anthropic, OpenAI in everyway.' One simply states 'it just works' and they never went back.

  • Cost efficiency shocked the market

    Users marvel that it reportedly cost only ~$5 million to develop yet erased $500 billion in Nvidia's market value, calling it a win for open source over oligarchic capital.

  • Detailed research paper

    One commenter calls the R1 paper 'the most detailed LLM sota paper since 2019,' noting western labs will likely reproduce its methods.

Surprising patterns

  • Some users say they trust Chinese data practices MORE than American ones, reasoning that Chinese services are explicit about collection while American firms try to hide it — 'being clear is better than not.'
  • Web search integration can partially bypass censorship: R1 evades sensitive topics on its own, but when web search is enabled, it will discuss findings from retrieved sources.
  • The censorship is dismissed by some as irrelevant for business/code use-cases — one user notes 'doesn't matter for business use-cases' — creating a split between developer and general-knowledge users.

WHO SHOULD SKIP IT

Anyone who needs an LLM as a general knowledge tool or search replacement for topics touching on Chinese geopolitics, human rights, or history.

8.5/10 GYIBB verdict
Full review → Buy on Amazon →

Synthesised from 325 real owner comments across 5 platforms. Every point is grounded in the comments — no marketing, no AI guessing. How we do it →