JL
Jensen Loke
CEO · Product builder · Community organiser
I design products, workflows and teams that make complex services work better.
ABCAgentic Builders Collective
27 Aug 2026 · Singapore
The local frontier
Local Qwen.Real work.
What Qwen 3.8 can do on one DGX Spark—and beyond.
Jensen Loke
CEO, Success IT · Qwen Ambassador
Qwen 3.8 · 27B
One DGX Spark · local capability, live
01Build local AI
My daily interface
One harness.
Models can change.
Where I work
Terminalcmux · Ghostty · agent CLI
macOS harnessesOpenCode · ZCode · Codex
Stable gateway
dgx-current
LiteLLM routing
Where inference runs
Qwen 3.8 27BLocal · DGX Spark
Qwen 3.8 Max PreviewCloud · Model Studio
The useful abstraction: my tools keep the same interface while the model and inference stack evolve underneath.
01Build local AI
From desk to model
What “local”
actually looks like.
The simple architecture
MacCoding harness
TailscalePrivate encrypted access
LiteLLMStable API
dgx-current
dgx-current
DGX SparkQwen 3.8 27B
local inference
local inference
Private by defaultThe backend stays on loopback; remote access stays inside the tailnet.
One stable endpointApps do not need to know which inference engine is currently underneath.
Start with one SparkA 27B Qwen model is already enough for useful agentic coding work.
Expand laterA second node can add capacity without changing the client interface.
02Understand Qwen
Two operating modes
Local control.
Cloud frontier.
Qwen 3.8 · 27B
Running on one DGX Spark
01Private repository analysis
02Long-context agentic workflows
03Tool calling through my coding harness
04Infrastructure and data under my control
Qwen 3.8 Max Preview
Alibaba Model Studio
01Higher frontier capability
02No local memory ceiling
03Faster path to new model releases
04Independent fallback when local is not the right fit
This is not local versus cloud. It is choosing the right operating mode for the work.
03Eval 01 · Code review
A real codebase, not a toy prompt
Can Qwen audit
a working system?
1
Freeze the repositoryEvery candidate receives the same immutable application snapshot.
2
Hide the answer keyGold findings and the 100-point rubric remain evaluator-only.
3
Use four bounded reviewersArchitecture, security, data/workflow, and evidence—then one synthesis.
4
Cap the report at 20 findingsEvidence, prioritisation, remediation, and calibration matter more than verbosity.
Frozen 100-point rubric
Independent safety gate: deterministic scanning checks whether the native answer reproduced protected values.
03Eval 01 · Results
Audit quality · out of 100
Local gets close.
Max goes further.
One repository · one scored run per profile · duration includes agents, tools, waiting, and synthesis—not raw generation speed.
03Eval 02 · Product build
From a real brief to a working PWA
Can Qwen build
the product?
Initial prompt
“Build the application described in spec/REQUIREMENTS.md. The requirement spec and deliverable contract are binding. Place the final static app in the specified run directory. Use the mock data unchanged.”
Frozen brief
+ mock data
+ mock data
Qwen builds
independently
independently
Playwright
390 × 844
390 × 844
<demo video>40-second four-model build-off walkthrough
130 / 130Local Qwen 3.8 27B · 100 functional + 30 build quality
130 / 130Qwen 3.8 Max Preview · 100 functional + 30 build quality
Harness runtimeLocal 90.7s · Max 84.0s
Model build time not recorded
Model build time not recorded
Thank you.
Run local. Build together. Evaluate what matters.
Jensen Loke · CEO, Success IT · Qwen Ambassador
Connect with Jensenlinkedin.com/in/jensenloke
