Stable Code 3B
Stable Code 3B is a compact base model for code generation and completion. It was pretrained on 1.3 trillion tokens of text and code, with support for filling missing code between a prefix and suffix.
Model design and access
The tokenizer includes special tokens for fill-in-the-middle tasks. The model uses 32 decoder layers and a modified GPT-NeoX tokenizer. Code completion and conversational instruction following are separate behaviors; the Instruct release adapts this base for requests and dialogue.Stable Code 3B official model card
Practical use
Practical guidance. The base fits editor-like completion and controlled code-adaptation work. Include the relevant surrounding code and keep generated changes small enough to review. The published context limit is a capacity figure, not evidence that every long repository fragment will be used accurately.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | stable-code-3b |
| Context window | 16,384 tokens Published checkpoint sequence length; prompt and generated text share the window. Stable Code 3B official model card |
| Parameters | 2,796,431,360 in the architecture table.Stable Code 3B official model card |
| Training scope | 18 programming languages.Stable Code 3B official model card |
| License | Stability AI Community License.Stable Code 3B official model card |
Related models
- Stable Code Instruct 3B
Stability AI ยท Stable Code
Stable Code Instruct 3B adapts Stable Code 3B for software-development conversations and SQL generation.
Launch date not confirmedDownloadable