Strata on one RTX 3060: testing and Hermes setup

This Strata local AI test uses a single RTX 3060 12GB, 64GB of system RAM and an E5-2696 V4 CPU. The speaker reports an average decode speed of 49.2 tokens per second for a task that produced about 32K output tokens after 617 seconds of thinking. He attributes faster stretches of repetitive output to speculative decoding and reports peak system power around 350 watts. These are results from the demonstrated workload, rather than a general performance guarantee.

The vision test asks Strata to transcribe a handwritten cake recipe. The speaker checks that the result preserves a crossed-out oven temperature and its replacement. He then connects Hermes to Strata through a custom HTTP endpoint on port 9877 with a /v1 suffix. He advises keeping the endpoint on an internal network and reducing reasoning from extra high to medium when using the agent harness.

The demonstrated configuration uses a 64K context window and a default batch size of one. The speaker says effective sub-agent use would require batch two or four and more memory resources; he also warns that context compaction can take a long time. Hermes subsequently pulls a repository and deploys a CT viewer. The hardware breakdown puts the demonstrated build at $825, or $769 with an E5-2695 V4 instead. Those are the speaker's quoted build costs, not current price guarantees.