Nekomata 14B
Nekomata 14B continues pretraining Qwen-14B on about 66B tokens of Japanese and English text.
Research release
The larger release uses a different training system
rinna trained this release on 16 AWS Trainium trn1.32xlarge nodes in about seven days. Its sequence limit is 8,192 tokens, shorter than the 7B Nekomata release.Nekomata 14B official model card
Do not infer context length from parameter count
Practical guidance. The larger model is not automatically the choice for longer input. Compare answer quality and usable sequence length as separate properties before selecting 7B or 14B.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | nekomata-14bnekomata-14b-gguf |
| Context window | 8,192 tokens Maximum sequence length stated in the model card. Nekomata 14B official model card |
| License | Tongyi Qianwen License Agreement; this is not an Apache-licensed Qwen2 release.Nekomata 14B official model card |
| Ancestry | Qwen/Qwen-14B.Nekomata 14B official model card |
Related models
- Nekomata 14B Instruction
rinna · Nekomata
This model instruction-tunes Nekomata 14B, retaining its Qwen ancestry and using the Alpaca task format.
Launch date not confirmedDownloadable - Nekomata 7B
rinna · Nekomata
Nekomata 7B continues pretraining Qwen-7B on about 30B tokens of Japanese and English text.
Launch date not confirmedResearch releaseDownloadable