Tensor TownISSUE #108 of 120

open weights vs closed · the divide

NeuraVSThe Overfit Ogre

Neura saysOpen weights vs closed models is a core divide — control and privacy vs convenience and frontier quality.

Open-weight models (Llama, Mistral, Qwen) can be downloaded, run privately, fine-tuned, and audited — control, privacy, no per-token fee, but you handle infra and they often trail the absolute frontier. Closed models (Claude, GPT, Gemini) offer top capability via API with no ops burden, but you send data out, pay per token, and can't inspect or self-host. Choose by privacy, cost, control, and capability needs.

Power-ups you unlock

Open: run private, fine-tune, audit, self-host
Closed: top capability via API, no ops
Open = control/privacy; closed = convenience/frontier
Choose by privacy, cost, control, capability

The Overfit Ogre attacks — common mistakes

Sending sensitive data to a closed API when open fits
Assuming open weights match the frontier
Underestimating self-hosting ops cost

Boss battlePick open vs closed for a privacy-sensitive on-prem app.

Example code

<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>open (Llama): private, self-host, fine-tune, audit
closed (Claude/GPT): frontier quality via API
privacy-sensitive on-prem → lean open</pre></body></html>

▶ Open the interactive comic issue

‹ Claude · Gpt · Gemini · Llama · The Lineup Local Inference · Ollama · Llama.Cpp · Lmstudio ›