Meet the team that turns your ideas into work.

Discover EightBots
EIGHTBOTS
Contact salesSign up
rinnaBakeneko

DeepSeek R1 Distill Qwen2.5 Bakeneko 32B

This Bakeneko release adds a chat vector from DeepSeek-R1-Distill-Qwen-32B to rinna's Japanese-adapted Qwen2.5 base, then applies ORPO.

Distilled reasoning enters through a merge

The documented merge uses the DeepSeek-distilled model minus Qwen2.5-32B with weight 1.0. The final model follows the DeepSeek-R1 chat format and tokenizer conventions.DeepSeek R1 Distill Qwen2.5 Bakeneko 32B official model card

Credit both the adapted base and the reasoning source

Practical guidance. This is rinna's trained derivative, with important upstream contributions from Qwen and DeepSeek. Evaluate Japanese reasoning directly instead of assuming that a distilled-model name transfers every capability of the full R1 model.

Specifications & access

DeepSeek R1 Distill Qwen2.5 Bakeneko 32B specifications
DetailValue and conditions
AccessDownloadable
Model identifiersdeepseek-r1-distill-qwen2.5-bakeneko-32bdeepseek-r1-distill-qwen2.5-bakeneko-32b-awqdeepseek-r1-distill-qwen2.5-bakeneko-32b-ggufdeepseek-r1-distill-qwen2.5-bakeneko-32b-gptq-int4deepseek-r1-distill-qwen2.5-bakeneko-32b-gptq-int8
LicenseApache 2.0 is declared for rinna's release.DeepSeek R1 Distill Qwen2.5 Bakeneko 32B official model card
Training dataORPO uses about 1,200 curated samples generated by DeepSeek-R1.DeepSeek R1 Distill Qwen2.5 Bakeneko 32B official model card

Related models

  • Bakeneko 32B continues pretraining Qwen2.5-32B on about 18B tokens from Japanese and English corpora.

    Launch date not confirmedResearch releaseDownloadable

  • QwQ Bakeneko 32B combines the Japanese-adapted Bakeneko base with a reasoning chat vector from QwQ-32B, followed by ORPO.

    Launch date not confirmedDownloadable