Hermes + Qwen 3.6 + NVIDIA DGX Spark: The Local AI Convergence

The Stack

• Hardware: NVIDIA DGX Spark (128GB unified memory, 1 petaflop) — purpose-built for 24/7 local agent operation.
• Models: Qwen 3.6 (Alibaba) — strong open-weight performance with low VRAM requirements.
• Agent framework: Hermes Agent — self-evolving skills, 3-layer memory, OpenShell sandbox.
• Key architecture: a self-evolving loop with full local execution and no external data sharing.

Software Summary

Qwen 3.6 Model

• 35B full precision or 27B quantized variant
• Runs efficiently on ~20GB VRAM
• Excellent reasoning and tool-use capabilities
• Fully open-weight (Alibaba)

Hermes Agent Framework

• 3-layer persistent memory system
• Self-evolving loop without human intervention
• NVIDIA OpenShell sandbox for safe execution
• Skills auto-improve over time

Hardware Summary

Setup Tutorial

01 — Hardware & Base Environment

• Confirm DGX Spark provisioned with 128GB unified memory.
• Install latest NVIDIA drivers + CUDA 12.4+ toolkit.
• Set up a dedicated user + persistent storage volume.

nvidia-smi  # verify 128GB and CUDA version

02 — Install Hermes Agent Core

• Install from the public Hermes / NousResearch distribution you use (or your internal fork). Upstream packages move; prefer the project docs over a hard-coded path.
• Create an isolated Python environment and install dependencies from that release.

Example shape only — replace the clone URL with the release you trust (public Hermès agent repos under NousResearch, or your private mirror). Do not paste unreviewed third-party install scripts on a DGX with production keys.

# Example — pin to a known release / commit
git clone https://github.com/NousResearch/Hermes-Agent.git
cd Hermes-Agent && uv venv && source .venv/bin/activate
# follow upstream install docs for the release you checked out

03 — Configure the 3-Layer Memory System

• Enable short-term, long-term, and skills memory in the agent config for your install.

Field names vary by Hermes version. Treat the block below as a conceptual checklist, not a drop-in config file.

# Conceptual example — map to your config schema
memory:
  short_term: true
  long_term: true
  skills: true

04 — Deploy the Qwen 3.6 Model

• Choose a variant: 35B (full) or 27B (quantized).
• Use an Ollama or llama.cpp backend.

ollama run qwen3.6:27b-q4

05 — Enable the NVIDIA OpenShell Sandbox

• Restrict execution to the sandbox only.

tools:
  open_shell: enabled
  sandbox_mode: strict

06 — Activate the Self-Evolving Loop

• Enable the core loop in the agent runtime.

loop:
  enabled: true
  stages: [observe, reflect, tool_use, code_gen, deploy]

07 — Initial Skill Seeding & Testing

• Seed with 3–5 core skills.
• Run a controlled test task.
• Monitor memory growth and loop iterations.

08 — Security & OpSec Hardening

• Disable all external telemetry.
• Set strict firewall rules.
• Verify zero data leakage.

Verification Checklist

• Model loads within target VRAM
• 3 memory layers persist across restarts
• Self-evolving loop produces an improved skill within 24h
• No outbound connections except intentional tool use
• Agent safely executes code inside the OpenShell sandbox