Start with the exact model and license, not a family name or an open-source label. Downloadable weights alone do not establish all the freedoms or information required by the Open Source AI Definition.
Read the exact model card
The checked examples include Llama 3.3 70B Instruct under its own license, Qwen3-8B under Apache 2.0 and DeepSeek-R1 under MIT. These are concrete license examples, not a claim that they are the newest or best models available.
What to record
Save the model identifier and revision, license, intended use, limitations, tokenizer and runtime requirements. Check a derivative or quantized release separately; do not assume its publisher is the original model author.
Plan deployment
Estimate memory, context and concurrency for the actual runtime and precision. Run a representative workload and record latency, accepted output quality and operating cost. A generic GPU estimate cannot cover every quantization or serving setup.
A fair evaluation
Use a fixed set of prompts with answer criteria. Record failures and unsupported claims, and compare a self-hosted option with the full cost of operating it rather than only a download price.
Terminology
Use open weights when that is what has been established. Refer to the OSI definition before claiming that an entire AI system is open source.




