Mixtral
About
The Mixtral family of large language models (LLMs), developed by Mistral AI, offers a groundbreaking approach in open-source AI through a sparse mixture-of-experts (SMoE) architecture. This innovative design allows the models to manage a significant number of parameters while ensuring efficient inference speed by activating only a subset of parameters for each token. Such architecture enables Mixtral models to deliver performance on par with much larger models, standing out in various benchmarks and outperforming competitors like Llama 2, and even equaling the prowess of closed-source models such as GPT-3.5. These models are multilingual, supporting languages such as English, French, Italian, German, and Spanish, and excel in domains like code generation. Instruction-tuned versions like Mixtral-8x7B-Instruct-v0.1 cater to applications requiring robust instruction-following and chat capabilities. The Mixtral family provides versatile models of differing sizes, addressing diverse computational and application requirements.
Specifications(5 models)
| Model | Released | Context | Parameters | Fn Calling |
|---|---|---|---|---|
| Mixtral 8x22B Instruct v0.3 | 2024-07 | 64K | 8x22B | Yes |
| Mixtral 8x22B v0.1 | 2024-04 | 64K | 8x22B | No |
| Mixtral 8x22B Instruct v0.1 | 2024-04 | 64K | 8x22B | No |
| Mixtral 8x7B | 2023-12 | 32K | 8x7B | No |
| Vultr Mixtral 8x7B | 2023-12 | 32768 | 47000000000 | No |
Available From(20 providers)
Pricing
Frequently Asked Questions
- What is Mixtral?
- The Mixtral family of large language models (LLMs), developed by Mistral AI, offers a groundbreaking approach in open-source AI through a sparse mixture-of-experts (SMoE) architecture. This innovative design allows the models to manage a significant number of parameters while ensuring efficient inference speed by activating only a subset of parameters for each token. Such architecture enables Mixtral models to deliver performance on par with much larger models, standing out in various benchmarks and outperforming competitors like Llama 2, and even equaling the prowess of closed-source models such as GPT-3.5. These models are multilingual, supporting languages such as English, French, Italian, German, and Spanish, and excel in domains like code generation. Instruction-tuned versions like Mixtral-8x7B-Instruct-v0.1 cater to applications requiring robust instruction-following and chat capabilities. The Mixtral family provides versatile models of differing sizes, addressing diverse computational and application requirements.
- How many models are in the Mixtral family?
- The Mixtral family contains 5 models.
- What is the latest Mixtral model?
- The latest model is Mixtral 8x22B Instruct v0.3, released in 2024-07.
- How much does Mixtral cost?
- Mixtral models range from $0.18/1M to $54/1M input tokens depending on the model and provider.