Define what you need to control
Start with the job. State whether you need local operation, an adaptable model, access to its weights or a particular hosting arrangement. Record the reason for each requirement so that a deployment choice can be tested against it.
Use this guide to record whether you can run, adapt and inspect a model for the intended task. It includes no ranked model list or measured hardware recommendation.
Check the exact release and licence
Record the model's exact release, source and licence. Check the rights required for the intended use, including modification and redistribution where relevant. Keep those findings attached to the release you assessed.
Check model access and service terms separately. Obtaining model weights and using a hosted service involve different evidence. A familiar model name alone does not establish the terms of the service that runs it.
Test the complete deployment
Write down the runtime, hardware and model configuration used in the test. Record memory needs, context settings and any changes to the supplied model. Run a fixed sample of the actual work on that setup.
Check the workflow's required tools and output formats. Inspect failures, response time and the effort needed to review the result. Keep the original task, accepted result and test configuration together so that another run can reproduce the comparison.
Compare accepted work
Use the same task and acceptance criteria for each candidate. Record which results pass and why others fail. Treat external benchmark scores as evidence about their stated evaluation. Your task still needs its own acceptance check.
Include operation and review in the cost record. Measure use at the system boundary you will pay for. Keep model capability, runtime behaviour and hosting terms as separate findings.
The Token Procurement guide explains how to join those findings to a purchasing decision. The small language models guide uses the same method for a narrower task.
Keep the evidence gap visible
This local guide covers licence, hardware and quality questions. The supplied selector evidence has no exact weekly Arena cost binding for open-weight routes. No measured capacity or quality score is assigned here.
Before choosing a deployment, complete the exact release, licence, hardware and task record. The AI in China page links to the named sources for that subject.