dQwen3.5 Release - MBPP+ fence: leaderboard
Base diffusion-model checkpoints; AR comparators follow the exact original publisher paths in the source manifest, including post-trained unsuffixed Qwen3.5 artifacts; 50B/100B adaptation tokens are distinct SWA trained checkpoints, never effort switches. Full-canvas/block decoding are separate boards; one token per step. MMLU5-shot likelihood; GSM8K8-shot1319; MATH5004-shot; HumanEval0-shot164; MBPP3-shot500; MBPP+3-shot378. No fewshot-dlms resampling fusion. Plain Python fence prompting; stock scores duplicate full-canvas tables, deltas are diagnostics. Metric: Accuracy (%). Higher is better. Source: https://arxiv.org/html/2609.20751v1
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Dream-Coder-7B (Base) | 64.81 | |
| 2 | dQwen3.5-9B-Base (100B adaptation tokens) | 59.79 | |
| 3 | Dream-7B (Base) | 59.52 | |
| 4 | dQwen3.5-9B-Base (50B adaptation tokens) | 58.2 | |
| 5 | dQwen3.5-4B-Base (50B adaptation tokens) | 53.7 | |
| 6 | dQwen3.5-4B-Base (100B adaptation tokens) | 52.38 | |
| 7 | dQwen3-1.7B-Base (100B adaptation tokens) | 49.47 | |
| 8 | dQwen3.5-2B-Base (100B adaptation tokens) | 44.71 | |
| 9 | LLaDA-8B-Base | 44.71 | |
| 10 | CoDA (Base) | 43.65 |
Interactive version: theaggregate.ai/benchmark?slug=dqwen3-5-release-mbpp-plus-fence · How It Works · Data refreshed daily, snapshot 2026-10-17.