Article 1 ·
My Local-LLM Code Review Benchmark Was 73% Ghosts
I benchmarked local LLMs as code reviewers on a DGX Spark, then audited the ground truth: 73 of 99 bugs were fixed before any reviewer saw them. What survived.
Spark Bench
What one commodity AI box, a DGX Spark on a desk, can and cannot be trusted to do, measured honestly. Local models as code reviewers and coding agents, two models served from one box with vLLM, model selection by measurement, and the evals behind every claim, failures included.
Article 2 ·
Two Local LLMs Serve My Coding Agents From One DGX Spark
Two local LLMs share one DGX Spark's memory for free, but running both at once costs the fast model 64% of its speed and hides four failure classes.
Article 3 ·
Two Local LLMs Replaced One Big One as My Coding Agent
Nemotron-3.5-Lightning drives my agent loops and Qwen3.8-27B takes code and vision on one DGX Spark; the 80B driver is retired. Every row measured.
Article 4 ·
Qwen3.8 Beat My Driver on Every Test. Nemotron Got the Job.
Qwen3.8-27B beat my DGX Spark's driver on coding, vision, and tools, then lost the Hermes driver seat to Nemotron on decode speed: 78.8 vs 11.3 tok/s.
Article 5 ·
My DGX Spark Reads 96% Busy and Is Mostly Waiting
nvidia-smi read my DGX Spark 96% busy; Nsight showed 3% issuing. Eight Qwen3.8 streams cut review wall time 4.7x; Nemotron beside them cost 54%.
Article 6 ·
The Review Bot Read 28 of My PRs and Caught 3 of 35 Bugs That Shipped
CodeRabbit flagged 3 of 35 bugs that shipped. A local LLM finder plus a Claude Sonnet 5 triage keeps 12 of its 19 at $0.48 a PR, so that runs on every push.