Qwen2.5 Bakeneko 32B
Bakeneko 32B continues pretraining Qwen2.5-32B on about 18B tokens from Japanese and English corpora.
Research release
One adapted base supports several later releases
The model uses the original Qwen2.5 tokenizer. rinna subsequently builds instruction and reasoning variants through different chat-vector merges and preference-training methods.Qwen2.5 Bakeneko 32B official model card
Use this checkpoint to isolate language adaptation
Practical guidance. Compare the base with Qwen2.5 on the same Japanese completion tasks before testing chat variants. The later releases add several training operations, so their gains cannot all be assigned to continued pretraining.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | qwen2.5-bakeneko-32b |
| Architecture | 64 layers with hidden width 5,120.Qwen2.5 Bakeneko 32B official model card |
| Ancestry | Qwen/Qwen2.5-32B.Qwen2.5 Bakeneko 32B official model card |
| License | Apache 2.0.Qwen2.5 Bakeneko 32B official model card |
Related models
- Qwen2.5 Bakeneko 32B Instruct
rinna · Bakeneko
Bakeneko Instruct combines the Japanese-adapted base with Qwen2.5's instruction chat vector and Simple Preference Optimization, or SimPO.
Launch date not confirmedDownloadable - QwQ Bakeneko 32B
rinna · Bakeneko
QwQ Bakeneko 32B combines the Japanese-adapted Bakeneko base with a reasoning chat vector from QwQ-32B, followed by ORPO.
Launch date not confirmedDownloadable