Three Qwen3.8 checkpoints benchmarked locally on RTX PRO 6000

On October 3, 2026, a post appeared on Reddit's LocalLLaMA community describing a local test of three Qwen3.8 checkpoints on a single RTX PRO 6000. According to the author, the comparison spanned three weeks and covered the same 10 tests. All three models received the same prompts and ran with the sampler settings listed on their own model cards, with one attempt per task. The results were published alongside a video of the runs. The author also notes that an earlier test had covered Qwen3.8-Flash-Next on its own.

The three checkpoints compared are RadixArk/Qwen3.8-27B-NVFP4 (dense), orcarouter/Qwen3.8-27B-Uncensored-NVFP4 (a dense, uncensored fine-tune), and RadixArk/Qwen3.8-Flash-Next-NVFP4 (MoE). The post's title frames the work as a comparison of new SGLang/vLLM recipes.

What the summary and excerpt do not provide is the substance of the individual tests, any scores, latency figures, or throughput numbers, nor any indication of which of the three checkpoints came out ahead. As it stands, the verifiable details are the test setup, the model list, and the conditions under which the runs were made. The account stops at the author publishing the write-up and the accompanying video; the benchmark conclusions themselves are not yet reflected in the coverage. Only this one report is available so far, so there are no competing figures from other sources to set against it.