Qwen3.8 Flash-Next hits 250+ tok/s on a single RTX 5080

On November 6, 2026, a user in the Reddit LocalLLaMA community posted a first-hand benchmark claiming that Qwen3.8 Flash-Next runs at over 250 tok/s in real Agent coding sessions on a single RTX 5080 paired with 64GB of system memory, with prefill reaching 5700 tok/s. The post attributes the gains to Basalt, the inference engine doing the heavy lifting. Comparing the same model file on the same machine against llama.cpp, the user reports roughly 3.5x faster decode without drafting, around 5x with the Lamina drafter attached, and about 7x faster prompt processing. Lamina is described as a further-trained MTP drafter from eddoursul.

The second part of the post concerns quantization. Unsloth's UD-IQ3_XXS checkpoint, uploaded on November 4 under the Dynamic v3.1 recipe, matches the footprint of the GSQ-RCO IQ3_S format commonly used on 16GB cards while being measurably more accurate: wikitext perplexity drops from 3.382 to 3.247, which the user notes essentially closes the gap to BF16 at 3.273. The trade-off is a 1 to 3.5% loss in decode speed.

As of now this is a single user report on a consumer-grade setup, with no independent reproduction or official figures to corroborate the speeds, the perplexity comparison, or the relative claims against llama.cpp.