Proprietary vs. Open-Weight vs. Open-Source: Decision Framework

The three-way classification of LLMs each represents a different business model and set of trade-offs.

Proprietary API (OpenAI, Anthropic, Google):

• Best for: fast prototyping, highest quality ceiling, no GPU management
• Limitations: data leaves your infrastructure, vendor lock-in, can't inspect weights
• Cost: $0.002-0.06/1K output tokens. At 100M tokens/day: $200-$6,000/day
• Choose when: speed-to-market matters most, data sensitivity is low, task complexity is high

Open-weight (Meta LLaMA 3.1, Mistral, Qwen, DeepSeek):

• Weights are downloadable; can self-host or run via managed cloud (Bedrock, Vertex)
• Can fine-tune within license terms (LLaMA: commercial OK, no redistribution for >700M MAU)
• Quality: LLaMA 3.1 70B ≈ GPT-4 quality on many tasks (as of late 2024)
• Self-host cost: A100 80GB GPU ≈ $3/hr; LLaMA 70B needs 2xA100 at ~16 tokens/s
• Effective cost at scale: $0.0003-0.002/1K tokens (vs API $0.01-0.03)
• Choose when: data privacy required, volume > 1M tokens/day, customization needed

Open-source (Apache/MIT licensed weights):

• GPT-2, BLOOM, Falcon, Mistral 7B (Apache), some Qwen variants
• No usage restrictions, can redistribute, modify, build services on top
• Quality is typically behind LLaMA 3.1 70B for complex tasks
• Choose when: need unrestricted redistribution, maximum legal clarity

The 1M tokens/day rule (Chip Huyen, Ch.4): Below 1M tokens/day: API cost is lower than self-hosting (amortized GPU costs are too high at low volume). Above 1M tokens/day: self-hosting becomes cheaper. Calculate exact crossover for your specific model and GPU.