rinnaBakeneko
QwQ Bakeneko 32B
QwQ Bakeneko 32B combines the Japanese-adapted Bakeneko base with a reasoning chat vector from QwQ-32B, followed by ORPO.
A smaller merge coefficient retains the adapted base
The formula adds 0.8 times QwQ-32B minus Qwen2.5-32B and excludes the embedding layer. The release follows QwQ's chat format and usage guidance.QwQ Bakeneko 32B official model card
Measure reasoning cost as well as answer quality
Practical guidance. Compare full reasoning runs rather than only final-answer length. A model that reaches the same answer through a longer internal generation can have a different latency and resource profile.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | qwq-bakeneko-32bqwq-bakeneko-32b-awqqwq-bakeneko-32b-ggufqwq-bakeneko-32b-gptq-int8qwq-bakeneko-32b-gptq-int4 |
| License | Apache 2.0.QwQ Bakeneko 32B official model card |
| Training data | About 1,300 curated DeepSeek-R1-generated samples are used for distillation and ORPO.QwQ Bakeneko 32B official model card |
Related models
A Japanese reasoning model combining Bakeneko with a DeepSeek-R1-Distill-Qwen-32B chat vector and ORPO.
Launch date not confirmedDownloadable- Qwen2.5 Bakeneko 32B Instruct V2
rinna · Bakeneko
A Japanese instruction model that merges Bakeneko with QwQ Bakeneko.
Launch date not confirmedDownloadable