Starling Alpha Models by Berkeley Artificial Intelligence Research (BAIR)
Last refreshed 2026-05-19. Next refresh: weekly.
Details
About
The Starling Alpha family of large language models (LLMs), developed by the Berkeley NEST team, includes models like Starling-LM-7B-alpha and Starling-LM-7B-beta. These models are fine-tuned versions of OpenChat 3.5, employing Reinforcement Learning from AI Feedback (RLAIF) 15. Leveraging the Nectar dataset and advanced reward training and policy tuning pipelines, the models excel in conversational AI, content generation, and question answering, achieving high scores on the MT Bench benchmark, with the beta version scoring 8.12 2. Available on Hugging Face and other platforms, these open-source models have restricted licenses for commercial use and competition with OpenAI 5.
Decision facts
- Best fit
- math-heavy prompts
- Capability starting point
- Starling LM 7B Alpha with 8k context
- Lowest tracked input
- Not tracked
- Closest related family
- MOSS-Audio
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs 8k context and 7B parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Starling LM 7B Alpha | Use when the workload needs 8k context and 7B parameters. | 2024-02 | 8k context7B parameters | Current |
Release Timeline
1 release groupSpecifications(1 models)
| Model | Released | Context | Parameters |
|---|---|---|---|
| Starling LM 7B Alpha | 2024-02 | 8k | 7B |
