Zephyr
About
The Zephyr family comprises advanced large language models specifically designed to function as highly responsive digital assistants. Notable for their human-like conversational abilities, these models excel in applications involving chatbots and virtual assistant roles. Developed with cutting-edge techniques such as distilled supervised fine-tuning (dSFT), AI feedback (AIF), and distilled direct preference optimization (dDPO), Zephyr models ensure that their output aligns closely with user intent. They often outperform larger models on certain benchmarks, despite being more compact 24. However, in areas requiring complex logic or specialized knowledge, they may face limitations 4. The Zephyr lineup includes iterations like Zephyr-7B-alpha and Zephyr-7B-beta, which is a fine-tuned variant of Mistral-7B 4513. Available through Hugging Face, these models are versatile tools for a range of natural language processing tasks 24.
Specifications(5 models)
| Model | Released | Context | Parameters | Structured Outputs |
|---|---|---|---|---|
| Zephyr 7B Alpha | 2023-10 | — | 7B | No |
| Zephyr 7B Beta | 2023-10 | — | 7B | No |
| Zephyr 7B Gemma | 2023-10 | — | 7B | No |
| Zephyr ORPO 141B | 2023-10 | — | 141B | Yes |
| Fireworks Zephyr-7B-beta | 2023-10 | 8K | 7B | No |
Available From(4 providers)
Pricing
| Model | Provider | Input / 1M | Output / 1M | Type |
|---|---|---|---|---|
| Zephyr 7B Beta | Replicate API | $0.05 | $0.25 | Serverless |
| Zephyr 7B Alpha | Replicate API | $0.05 | $0.25 | Serverless |
| Fireworks Zephyr-7B-beta | Fireworks AI | $0.1 | $0.1 | Serverless |
| Zephyr 7B Beta | Fireworks AI | $0.2 | $0.2 | Provisioned |
| Zephyr ORPO 141B | DeepInfra | $0.65 | $0.65 | Serverless |
Frequently Asked Questions
- What is Zephyr?
- The Zephyr family comprises advanced large language models specifically designed to function as highly responsive digital assistants. Notable for their human-like conversational abilities, these models excel in applications involving chatbots and virtual assistant roles. Developed with cutting-edge techniques such as distilled supervised fine-tuning (dSFT), AI feedback (AIF), and distilled direct preference optimization (dDPO), Zephyr models ensure that their output aligns closely with user intent. They often outperform larger models on certain benchmarks, despite being more compact 24. However, in areas requiring complex logic or specialized knowledge, they may face limitations 4. The Zephyr lineup includes iterations like Zephyr-7B-alpha and Zephyr-7B-beta, which is a fine-tuned variant of Mistral-7B 4513. Available through Hugging Face, these models are versatile tools for a range of natural language processing tasks 24.
- How many models are in the Zephyr family?
- The Zephyr family contains 5 models.
- What is the latest Zephyr model?
- The latest model is Zephyr 7B Alpha, released in 2023-10.
- How much does Zephyr cost?
- Zephyr models range from $0.05/1M to $0.65/1M input tokens depending on the model and provider.
