DEVNET AI LAB

DEVNET AI LAB — Home AI lab of DK

1,263 TOPS.

Four chip makers. Five machines. One house.

Scroll to open the stack

The stack

Five boxes. Each has one job.

Running total

  1. 1,000
  2. 1,126
  3. 1,225
  4. 1,263
TOPS
  1. 01NVIDIA

    DGX Spark

    Model host

    GB10 Grace Blackwell · 20 Arm cores · 128 GB unified

    vLLM serving Qwen 3.6 35B at ~54 tokens/s

    1,000 TOPS

  2. 02AMD

    Strix Halo

    Image & video

    Ryzen AI Max+ 395 · Radeon 8060S · 64 GB + 64 GB GPU

    ComfyUI: 357 images, 0 failures, median 32 s

    +126 TOPS

  3. 03Intel

    NUC

    Voice

    Core Ultra 9 285H · NPU + Arc · 64 GB

    Whisper speech-to-text, Kokoro text-to-speech

    +99 TOPS

  4. 04Apple

    Mac mini M4 Pro

    Control plane

    12 cores · 24 GB · Neural Engine

    LiteLLM gateway + AILogic

    +38 TOPS

  5. 05Apple

    Mac mini M6

    Dev server · home of LabScope

    12 cores · 24 GB

    Home of LabScope, Dmotion, Dpages

    Neural Engine not counted

Vendor peak figures.

Meet the machines

By the numbers

The whole lab, counted.

92

CPU cores

315

GB memory

251

GB GPU / unified memory

8

TB storage

21

models in the lab17 loaded now

74

cloud models from 6 providers

73

apps on the webbehind single sign-on by default

99.993%

uptime over 30 days

Read from LabScope, 26 Oct 2026.

Last 30 days

30 days through one gateway.

AI calls
10,635
tokens
62.4M
success
96.7%
median gateway overhead
5 ms

22.5 million tokens (36%) answered by models running in this house.

36% local64% cloud
Call mix · log scale
  • Chat9,169
  • Image generation772
  • Speech236
  • Web search94
  • Embeddings13

66 apps called 68 models. Today so far: 1,098 calls, 6.6 million tokens.

The failure test

We switched the gateway off.

Gateway upGateway off · 69 sRelays switch back

156 / 156

calls answered while the gateway was off

98

Gateway path

58

Backup path

69 seconds dark. 156 of 156 calls answered.

Every box runs a relay. Apps call it locally. It goes to the gateway first, and if the gateway is down it replays the request straight to the provider with its own backup key. 58 calls took the backup path. Every relay switched back by itself.

Earlier run: 72 s outage — 99 of 100 answered in 0.35–0.7 s.

How AI calls flow

What runs here

73 apps. A few of them.

Dmotion

AI animation and explainer-video studio

Mac mini M6

Dpages

Prompt to animated websites, decks, infographics

Mac mini M6

SparkChat

Private chat with an off-the-record mode

DGX Spark

Conductor

Multi-model agent orchestration

Intel NUC

Pilot

Browser agent

Intel NUC

OpenDots

Voice agent

Intel NUC

PacketLens

Network analysis

Intel NUC

SkillForge

Skills library

Mac mini M4 Pro

Also: Rushes · SketchMotion · Verdict · Books2Visuals · InfoFrame · Arch Explainer · webOS · Cbuse · imagin-ai · Isthatall · Dify · n8n · Flowise

Built by DK

LabScope watches all of it.

Agentless, read-only mission control. One SSH session per poll per host. Nothing installed, nothing written, never sudo.

  • Metrics every 15 s
  • 30 days of history
  • Prompts and answers never stored

Ask the lab by typing or speaking, and its fastest local model answers from the live snapshot.

Inside LabScope
LabScopeOverview5 hosts · 17 models loaded
  • Overview
  • Inference
  • Models fit check
  • Gateway story
  • Usage Sankey
  • Network
  • Capacity
  • Benchmarks
  • Fix list

DGX SparkQwen 3.6 35B · ~54 tok/s

Strix HaloQwen Image · median 32 s

Intel NUCWhisper · Kokoro

Mac mini M4 ProGateway · 5 ms median

Mac mini M6LabScope · 15 s poll

Ask the lab“Where can this run?”

Five machines. Four chip makers. 1,263 TOPS.

21 models at home, 74 in the cloud, one gateway that doesn’t go down.

And one tool that sees every machine, model and call.

DEVNET AI LAB — designed, built and run by DK.