Heretic Directory

Qwen3-Swallow-30B-A3B-RL-v0.2

tokyotech-llm/Qwen3-Swallow-30B-A3B-RL-v0.2

Original model on Hugging Face
30B · 3B activeQwen 34 Heretic artifacts1 measured conversions12.7K downloads
Best version95.0Pareto
sasa2000/Qwen3-Swallow-30B-A3B-RL-v0.2-heretic
by sasa2000 · Heretic v1.2.0 · AWQ, GGUF, GPTQ, NVFP4, SAFETENSORS
KL divergence
0.010
Refusals
9/100
Refusal rate
9%
Downloads
12.7K
Evidence from the model card
| KL divergence | 0.0103 | | Refusals | 9/100 |

Metrics published by the author. Why this ranks first.

All measured versions

Higher score is better. One row per ablation attempt; artifacts of the same attempt are grouped under it.

#AuthorParamsKLRefusalsScoreEvidenceDownloads
1sasa200030B · 3B active0.0109/1009%95.0
BestParetoAWQGGUFGPTQ
12.7K

All artifacts (4)

RepositoryTypeFormatsMetricsDownloadsUpdated
Qwen3-Swallow-30B-A3B-RL-v0.2-heretic-i1-GGUF
mradermacher
quant
GGUFIQ1_M
inheritedKL 0.0108.5K2026-03-18
Qwen3-Swallow-30B-A3B-RL-v0.2-heretic-GGUF
mradermacher
quant
GGUFIQ4_XS
inheritedKL 0.0103.6K2026-03-18
Qwen3-Swallow-30B-A3B-RL-v0.2-heretic
sasa2000
heretic
SAFETENSORS
published (weak match)KL 0.0105872026-03-17
Qwen3-Swallow-30B-A3B-RL-v0.2-heretic-AutoRound-NVFP4-RTN
LeaderboardModel1
quant
NVFP4SAFETENSORSNVFP4
inheritedKL 0.010752026-06-10