Skip to content
View marksunner's full-sized avatar

Block or report marksunner

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
marksunner/README.md

Mark Sunner

Former CTO (MessageLabs, acquired by Symantec 2008). Now retired, happy tinkering and building local AI infrastructure on NVIDIA DGX Spark hardware.

I work with AI agents daily and much of what you see here is built collaboratively. I provide the hardware, the direction, and the domain experience; the agents handle the heavy engineering. It's a genuine partnership and I'm not shy about it.


Local AI via DGX Spark Infrastructure

Running 122B–671B parameter models on single, dual and cluster NVIDIA DGX Spark systems. Real benchmarks, real hardware, no cloud.

dgx-spark — Start here. Navigational guide to all DGX Spark work, organised by inference engine and model. → dgx-spark-glm52 ✨ — Cluster deployment guides written from real experience: GLM 5.2 on 4× Sparks (the complete journey from unboxing to first inference) · What Is Fabric? (building a lossless RoCE fabric with the MikroTik CRS812, from first principles)

Inference engines


Intelligence pricing

  • barrel-indexThe Barrel Index: Pricing benchmarking for machine intelligence. A single HTML file that shows what every major AI producer charges per million tokens, pulled live from the market and racked like a commodity board. No build step, no dependencies, no API key.

Other interests

  • storytelling-psychology — Communication frameworks grounded in cognitive psychology. Built from years of helping technical teams make complex ideas land.
  • helix — A non-linear cognition interface for voice AI, designed for dyslexic thinking. Currently paused while pursuing infrastructure work.
  • decoherence-paper — Speculative physics: quantum decoherence and the emergence of classical reality. Worldbuilding for a hard SF novel, not a contribution to physics.

Pinned Loading

  1. atlas atlas Public

    Forked from Avarok-Cybersecurity/atlas

    Pure Rust Inference Engine

    Rust

  2. dgx-spark-single-stack dgx-spark-single-stack Public

    Complete single-Spark AI agent stack: Qwen 122B + Hermes + Honcho on one DGX Spark

    16

  3. dgx-spark-step37-flash dgx-spark-step37-flash Public

    Step 3.7 Flash (198B MoE) on a single DGX Spark — 27 tok/s, 128K context, agent-ready

  4. dgx-spark-vllm-tp-benchmark dgx-spark-vllm-tp-benchmark Public

    DeepSeek V4 Flash dual DGX Spark benchmark: vLLM tensor parallelism (TP=2) over NCCL/RoCE — 12.4 tok/s

  5. storytelling-psychology storytelling-psychology Public

    The psychology of making ideas land — communication frameworks, narrative transportation, and persuasion science

  6. decoherence-paper decoherence-paper Public

    Speculative physics paper on quantum decoherence and the emergence of classical reality