Published 2026-10-12

This test evaluates whether the Strata local AI engine runs on a single NVIDIA RTX 3060 with 12 GB of VRAM, rather than the two-GPU configuration used previously. The system also tests a second CPU option, the Intel Xeon E5-2695 v4, against the E5-2696 v4 used in the test machine.

Test Configuration

The E5-2696 v4 was not part of Intel's standard SKU lineup and does not appear on Intel's ARK product page.

Initial Setup and Analysis

Strata was started on the single-GPU system with the Vision stack enabled. As a first task, the engine was asked to review the Intel E5 v4 family page and the Strata repository. It was then asked to estimate the performance impact of switching from the E5-2696 v4 to the E5-2695 v4.

The E5-2695 v4 has 18 cores. The E5-2696 v4 has 22 cores.

Performance Results

Long-Context Reasoning

Speculative decoding began to engage above 16,000 tokens. The speaker attributed this to prompts in which only small changes occurred between requests.

Handwritten Note Transcription

The engine transcribed a handwritten recipe for banana chocolate chip cake. The handwriting was described as not fully legible.

The model correctly identified the following:

• A crossed-out oven temperature of 350
• The corrected temperature of 325
• A baking time of 30 to 35 minutes

Prefill speed for this task was expected to reach the 800s tokens per second. Decode speed was expected to be faster than in the reasoning test because the output was shorter.

CPU Comparison: E5-2695 v4 vs. E5-2696 v4

The E5-2696 v4 benefits from sustained all-core maximum performance. Whether the system BIOS allows sustained all-core operation determines whether this advantage applies.

The speaker stated that no large user-facing difference in tokens per second is expected between the two CPUs.

Hermes Agent Integration

Hermes agent can connect to Strata through a custom model endpoint. The procedure is as follows:

1. Open the custom model endpoint setting in the Hermes CLI or desktop application.
2. Enter the internal server URL. The connection uses HTTP rather than HTTPS.
3. Enter the port number. In this test, the port was 9877.
4. Append the path /v1.
5. When prompted for an API key, leave it blank. The endpoint runs on an internal network.
6. In API mode, accept the auto-detected model name.
7. Leave context window detection enabled.

The speaker recommended against exposing the endpoint to the public internet. For internal deployment, the speaker recommended running the service in Proxmox to limit the blast radius of any compromise.

After configuration, Hermes displayed the correct provider and communicated with Strata over the network.

Configuration Limits

• Context window: Set to 64K in this test. A 128K window produces long compaction cycles because of the slower decode speed.
• Sub-agents: Not supported in the default configuration because batch size defaults to one. Batch sizes of two or four are required.
• Hardware for sub-agents: Two GPUs and more system memory are required. The speaker stated that 128 GB of DDR4 RAM would allow sub-agents on this platform, but did not measure the performance impact.
• Context at 64K with two GPUs: The speaker expected this configuration to run at 64K.
• Reasoning level: The speaker recommended changing the reasoning setting from extra high to medium when using a harness such as Hermes.

The model was referred to as "strata 3060" in the Hermes configuration.

Application Demonstration: Radial CT Viewer

The speaker used the model to build a WebRTC viewer for radial CT imaging. The viewer includes controls for changing the level of detail. The demonstration included a soft tissue view.

The speaker stated that the viewer is not medical software and is not intended for medical use.

During the demonstration, the speaker reported the following:

• Hermes agent, vision recognition, and code base deployment all ran as expected.
• The repository was retrieved, handled, and deployed successfully.
• The prefill speed of the 800s tokens per second was observed during this work.

The speaker used a model named "Quinn 3.8.27b" to build the viewer, as stated in the source.

Cost Analysis

The source gives the following totals:

• Single-GPU build with E5-2696 v4: $825
• Single-GPU build with E5-2695 v4: $769
• Two-GPU build: About $1,100

The source gives two different prices for the E5-2695 v4, about $75 shipped and about$20. The totals are reported as stated. All components were sourced from eBay. The speaker reported that GPU prices were rising.

Findings

• A single RTX 3060 with 12 GB VRAM runs Strata with 64 GB of system RAM.
• Decode speeds in the mid-40s to 49.2 tokens per second were measured during long-context reasoning.
• Prefill speeds in the 800s tokens per second were observed.
• The E5-2695 v4 and E5-2696 v4 are not expected to differ significantly in user-facing token generation speed.
• Hermes agent works with Strata through a custom endpoint, with a 64K context window limit.
• Sub-agents require batch size two or four, which needs additional GPU and memory capacity.
• The total cost of the single-GPU build is $769 with the E5-2695 v4, or$825 with the E5-2696 v4.