The Best-Quality Local LLM for Each Scenario

After a model finishes a scenario, its output is graded from 0 to 100 by an LLM judge (currently openai/gpt-5.4-mini) against a rubric written for that scenario. Because every score comes from one judge model, treat them as a consistent relative signal between models rather than an absolute measure of quality.

Only models with at least three runs under the current scenario set are ranked, so no single lucky score can top a chart. That leaves 24 models across 178 runs, drawn from the 253 runs recorded since the scenarios were finalized on 13 April 2026.

No model wins everything. The mid-size Qwen3.6 family dominates planning and analysis, a large mixture-of-experts coding model takes roleplay, and code generation stays stubbornly hard for all of them.

The roleplay result is the surprise: qwen3-coder-next, a model built for code, produces the most immersive and in-character narrative in the data. The same model scores far lower on research, a reminder that strength in one scenario does not carry to the next.

If you would rather run one model for everything, Qwen3.6-27B at 4-bit is the clearest pick: a combined score of 83.2 across the four scenarios, topping or tying the field in three of them and never dropping below 75.

If you want quality and speed together, the Qwen3.6-35B-A3B mixture-of-experts build is the better trade: it gives up only about three combined points while running roughly three times faster, because it activates only a few billion parameters per token.

The same judge is far stricter on some scenarios than others, which is a useful signal about where local models genuinely hold up. Creative and planning work scores highest; one-shot code generation scores lowest by a wide margin.

The practical takeaway: if your use case is coding, weight the quality scores heavily and test on your own workloads before committing, because even the best local models here produce competent rather than flawless code in a single pass.

Every result here comes from consumer hardware — the machines contributors actually own. The most common GPU is a 24GB AMD Radeon RX 7900 XTX, followed by Apple Silicon Macs (M5 Max and M4 Pro, with up to 64GB of unified memory) and a handful of NVIDIA RTX 40 and 50-series cards.

That ceiling defines what got tested: almost every ranked model is between 3B and 36B parameters, the range that fits in about 24GB of VRAM or an Apple unified-memory budget. No datacenter GPUs and no large full-precision dense models appear in the data. Read these rankings as the best models you can run at home today — add bigger models on more powerful hardware and the standings would very likely shift, since larger models tend to win on quality, especially on code generation.