Sherkala 8B Chat
Sherkala is an 8B instruction model focused on Kazakh, with English, Russian and Turkish support. Inception, MBZUAI and Cerebras developed it from the Llama 3.1 8B architecture. The published repository has an acceptance gate.
Restricted access
Why tokenization matters
The developer expanded the vocabulary to reduce the number of token pieces per Kazakh word. Continued pretraining used 45.3 billion tokens. This is language adaptation of an existing architecture, unlike Jais 2's training from scratch.Sherkala official model card
Separate research suitability from commercial rights
Practical guidance. Sherkala is useful for Kazakh-language experiments and comparisons. Its noncommercial license is a material selection constraint for a business product. Evaluate language quality and deployment rights as separate questions.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | Llama-3.1-Sherkala-8B-Chat |
| Context window | 8,192 tokens Published context length Sherkala official model card |
| License | CC BY-NC-SA 4.0Sherkala official model card |
| Vocabulary adaptation | 25% expansion with frequent Kazakh tokensSherkala official model card |
Related models
- Nanda 10B Chat
Inception42 (G42) ยท Nanda
Hindi-and-English chat model based on Llama 3 architecture and trained with 65 billion Hindi tokens.
Launch date not confirmedRestricted accessDownloadable