Benchmarks · Qwen35-A3B
One model. Two quants.
Qwen3.6-35B-A3B, quantized to NVFP4 two different ways, served on the same DGX Spark and judged by the same judges on the same day. Speed uses the identical single-stream metric for both. This page is the head-to-head; the full leaderboards live on StoryBench and DharmaBench.
Story by story
The same five prompts, scored 1–10 by independent Claude judges. Tap a row for both judges' full rationales.
| Story | NVFP4 (NVIDIA) | NVFP4-Fast (unsloth) |
|---|
Worldview (DharmaBench)
Method & provenance
Also in the family
Finetunes of the same base model on the main leaderboard — not part of the duel.
| Config | Ctx | Overall | Tok/s |
|---|