Looks like the docs have a typo:

    Recommended context: 65,536 tokens (can be increased)

That should be recommended token output, as shown in the official docs as:

    Adequate Output Length: We recommend using an output length of 65,536 tokens for most queries, which is adequate for instruct models.