Why DeepSeek had to be open source

Security, Local Use & Practicalities

• Several commenters verify that DeepSeek R1 (or its distillations) runs with no external network traffic and can be used fully offline.
• Concerns remain about supply-chain risk (e.g., verifying safe serialization formats such as safetensors vs unsafe ones like pickle).
• Full R1 is described as ~650–700GB (fp16) with quantizations around 150GB; only distilled models (based on Llama/Qwen trained on R1 outputs) are practical on single GPUs and consumer hardware.

Is DeepSeek “Open Source”?

• Large subthread argues DeepSeek is not open source but “open weights” or freeware:
  • Missing: training code, training data, and low-level PTX/cluster tooling.
  • Weights are likened to binaries or bytecode: modifiable via fine‑tuning, but not reconstructible from source.
• Others counter that for LLMs, weights are the “preferred form for modification,” citing GPL/OSI language and practical constraints (no one can afford to retrain frontier models).
• Nuanced taxonomy is proposed:
  • open-source inference code
  • open weights
  • open pretraining recipe (code + data)
  • open fine‑tuning recipe (code + data)
• Licensing nuances: older DeepSeek-V3 weights have a custom, more restrictive license; R1 and R1-Zero weights are MIT-licensed.

Trust, Censorship & Geopolitics

• Strong skepticism toward a Chinese API; open weights and local deployment are seen as crucial for Western adoption.
• Evidence cited that DeepSeek censors topics sensitive to the Chinese state; similar concerns are raised about Western models on other geopolitical topics.
• Some suggest using Chinese and Western models to cross-check each other; others argue both sides propagate their own narratives.

Competition, Moats & Economics

• Debate over whether DeepSeek “dethrones” OpenAI or just narrows the gap temporarily.
• Some see DeepSeek as proof that frontier-level reasoning can be built for a few million dollars, eroding proprietary moats and pushing prices toward zero.
• Others argue large incumbents (especially Google) still have substantial moats: custom hardware, data pipelines, userbase, and monetization channels.
• Expectation from many participants: a mixed future where open(-weights) models commoditize baseline capabilities, while proprietary models continue at the bleeding edge and for integrated commercial offerings.

Reaction to the Article Itself

• Multiple commenters call the post clickbait/content marketing for Lago and object to the claim that DeepSeek “proves” an open-source future; they see it as one data point, not a proof.