A benchmark of five locally run LLM agents on a single RTX 5090 shows that a 35-billion-parameter mixture-of-experts (MoE) model outperformed its peers on a real-world coding task. The test, which added a Tag Manager to an existing admin panel without any cloud APIs, crowned Qwen 3.6 35B-A3B as the clear winner.

Why the test matters

Running large language models on personal hardware lets developers dodge API fees and data-privacy worries. Yet “local agents” remain a buzzword: can they actually edit files, call command-line tools, and ship production-ready code without a human hand-holding them? This hands-on comparison, stripped of cloud services, gives developers a concrete sense of where the technology stands.

The hardware and the task

All five models ran on the same workstation: an RTX 5090 GPU, a typical high-end consumer card, and no external services. The task was deliberately simple but representative – add a Tag Manager component to an already-built admin section. Success required the model to locate the correct source files, edit them, and verify that the new feature integrated without breaking existing functionality.

Model performance

  • Qwen 3.6 35B-A3B (MoE) – Finished the job autonomously, added sensible improvements that weren’t requested, and needed no post-run patches.
  • Qwen 3.6 27B (dense) – Completed the task but took roughly twice as many interaction steps and occasionally mis-interpreted instructions.
  • GLM-4.7-Flash (dense) – Produced a working implementation but used incorrect file-matching patterns and omitted security checks, leaving the code vulnerable.
  • Qwythos-9B – Did not execute any real tools; the failure could stem from the model itself or from the local setup, but the test could not isolate the cause.
  • Nemotron-3-Nano (hybrid) – Got stuck in a loop, spending 40 turns searching for a folder it had already located, and never progressed beyond that point.

What the results reveal

Architecture is not a reliable predictor

The MoE model, which splits its parameters across multiple expert sub-networks, won outright, while the dense and hybrid variants split between success and outright failure. This suggests that raw architectural choices do not guarantee tool-use competence.

Tool use remains a major hurdle

None of the five agents used the dedicated screenshot tool supplied for the test. All tried to improvise, either by guessing file names or by attempting indirect workarounds. The gap between “can generate code” and “can orchestrate external utilities” is still wide.

Running locally means debugging the stack, not just the model

Two models required on-the-fly patches to their prompt templates before the test could even start. The effort spent fixing model-specific bugs eclipsed the time spent writing the actual Tag Manager code, highlighting how fragile current local deployments are.

A note of caution

The benchmark reflects a single hardware configuration, a single coding scenario, and a small suite of models. The Qwen 3.6 35B-A3B had previously flopped in an earlier round; its earlier failure was a fluke. The results are therefore indicative, not definitive.

What to watch next

Future rounds will need to expand the task set, include more diverse toolchains, and test on a broader range of hardware. Observers should track whether MoE models consistently outperform dense and hybrid designs, and whether developers can build reliable wrappers that eliminate the need for manual bug-patching.

Takeaway: A 35B MoE model can already act as a competent local coding assistant, but the broader ecosystem—tool integration, prompt engineering, and stable runtimes—still lags behind. Until those pieces click, developers should temper expectations about “plug-and-play” local agents.