Leitfaden
Open Weight vs Open Source AI Models
The difference between open-weight, open-source and proprietary AI models — and what each license actually allows.
The phrases "open source AI" and "open weights" are often used interchangeably in marketing copy, but they describe different things and license differently. Picking the wrong category for your use case can mean either over-paying for an API or breaching license terms when you ship.
Three tiers of openness
From least to most open:
- Proprietary — only an API exists. Weights, training data and source code are all private. Examples: GPT-5, Claude, Gemini.
- Open weight — trained model parameters are downloadable; you can self-host, fine-tune and run offline. Training data and training code are typically not released. Examples: Llama 3 family, DeepSeek V3 / R1, Qwen3, Kimi K2.
- Open source (in the OSI sense) — weights, training data and training code are all released under an OSI-approved license. Almost no frontier-class LLM meets this bar today.
License gotchas
Open-weight licenses are not all equal. Some common restrictions:
- Acceptable use clauses — most prohibit certain harmful applications (weapons, surveillance, etc.).
- Scale gates — Llama's license restricts free commercial use for products with >700M MAU.
- Distillation clauses — some licenses forbid using model output to train competing models.
When to choose open weights
Pick open weights when one of these is true:
- You need on-prem deployment for compliance (HIPAA, financial, defense).
- You want to fine-tune on private data without sharing it with a vendor.
- Cost at scale is dominated by inference, and self-hosting on GPUs you already own beats per-token API pricing.
- You need offline / edge inference (mobile, embedded, air-gapped).
Frequently asked questions
Are 'open-weight' and 'open-source' the same thing?
No. Open-weight means the trained weights are downloadable; open-source additionally means training data and code are released. Most models marketed as 'open' (Llama, Qwen, DeepSeek) are open-weight, not open-source.
Can I use open-weight models commercially?
Usually yes, but check the licence. Llama, Qwen and DeepSeek allow commercial use with caveats (e.g. attribution, MAU limits). Some research-only releases are not commercial-friendly.
Is self-hosting cheaper than using an API host?
Only at very high volume. A typical 70B open-weight model needs ~140GB of VRAM (2× A100s or 1× H100), which costs $4–10/hr on cloud GPUs. Below ~10M tokens/day, an API host is usually cheaper.