For months, anyone trying to replace a proprietary model with a self-hosted LLM has hit the same wall: benchmarks promised, but real work disappointed. A story posted on Reddit by a user running Qwen 3.8 Flash Next on two RTX 3090s adds a new piece to this puzzle, from an unexpected angle: not a new model, but a different harness.

The author had long used local models only for demos like generating a multi-level 3D Mario-style game. On serious work, no self-hosted LLM could keep up with GPT-5.2 or GPT-5.6 Luna, the latter described as worse in benchmarks. Then he asked GPT-5.6 Luna to configure Codex CLI for the local model. He had already tried pi.dev and opencode with unconvincing results. With Codex CLI, the first test on the 3D game prompt was the best he had ever seen. The biggest jump came on a real project: the same one Luna had been working on for days, given to both in parallel. The local model outperformed Luna, while Qwen and DeepSeek had previously failed on the same project. For web search, the user also ported the pi-smart-web-search skill into Codex CLI and found it worked extremely well.

The story is interesting not because of the model, but because of the independent variable: the harness. Codex CLI does not change the model's weights, but it changes how the model receives context, files, tools, and instructions. If a local LLM with the same base fails with pi.dev and opencode but excels with Codex CLI, the difference lies in integration, not in raw generative capability. This shifts attention from raw benchmark scores to context engineering and tool orchestration.

For anyone evaluating on-prem deployment, the message is structural: the bottleneck can be the harness, not VRAM or quantization. A machine with two RTX 3090s can be adequate if the workflow is well orchestrated, while a larger model in a mediocre harness may seem inadequate. From a hardware standpoint, two RTX 3090s are a common setup for local inference but often considered entry-level for mid-sized models. The story provides no throughput or latency figures, but it suggests that a context-aware harness can compensate for memory and context window limits. It is not a raw compute issue, but how the system handles the pipeline: file retrieval, tool calls, response formatting. Anyone designing local stacks should give harness selection the same attention reserved for quantization or model choice.

There is a second layer: harness choice becomes a factor in TCO and data sovereignty. If an open-source CLI unlocks acceptable performance on local hardware, the case for sending code to cloud APIs weakens. It does not disappear, but the calculation changes: per-token marginal cost gives way to fixed hardware cost and integration maintenance. The source itself reveals a tension: the user relied on a proprietary model to configure the tool that later made the local model competitive. That detail shows how much the self-hosted ecosystem still depends on fragmented knowledge and tooling. The final question on Reddit, whether an extension could bring pi.dev to the same level as Codex CLI, points to a documentation and standards gap.

The perceived difference between harnesses will continue to outweigh model benchmarks until these configurations are documented and reproducible.