DeepSeek-R1 has downloadable model weights as well as hosted options. Run a supported checkpoint locally when you need to control where inference happens, or use a hosted endpoint when you want someone else to run the model. Check which checkpoint and service you are actually using before comparing answers.
Choose a route
| Route | What to check |
|---|---|
| Provider's chat or API | Check the current model ID, provider terms, data handling and rate limits. A familiar product name does not pin a checkpoint. |
| Local application or runtime | Check that the runtime supports the intended checkpoint, quantization and hardware. Confirm that inference is local and review whether the app separately sends telemetry or uses hosted tools. |
| Self-hosted endpoint | Pin the weights and serving version. Plan for access control, logs, upgrades and resource limits before exposing it to other users. |
Do a reproducible first run
- Start from the publisher's model card. Check the exact base family, parameter count and license for the chosen checkpoint. A distilled model has a different size and may answer differently from the full model.
- Choose a small supported checkpoint and use your runner's memory estimate. Record the checkpoint, quantization and context setting.
- Try a harmless fixed question such as Why does a 1 TB drive show roughly 931 GiB? Check whether the model gives the conversion, and compare the result with the arithmetic.
- If the goal is offline inference, test again after the model has downloaded and networking is unavailable. That checks the tested inference path. Review the app's settings and logs separately before relying on its privacy properties.
Local hosting changes who runs inference and which settings you control. Model behavior still depends on the weights, prompt, runtime and any surrounding application rules. Keep those details alongside a comparison so someone else can reproduce it.