The world of artificial intelligence is pushing boundaries with the latest addition, Mixtral 8x7B. It is a revolutionary open-weight model with unique architecture, offering impressive performance and accessibility.
Most importantly, it focuses on making the community benefit from original models to foster new inventions and uses.
In this article, we will discuss the large language model, its capabilities and strengths, and how you can access this exciting technology.
What is the Mistral 8x7B?
Mixtral 8x7B is Mistral AI’s latest innovation. The “8x7B” refers to its structure: 8 groups of 7 billion parameter experts, resulting in a total of 56 billion parameters. It is a high-quality sparse mixture of expert models (SMoE) with open weights. According to Mistral AI’s official blog, it outperforms Llama 2 70B on most benchmarks, with 6x faster inference. It is seen as the strongest open-weight model with a permissive license and the best model overall regarding cost/performance trade-offs.
Mixtral Capabilities
Unlike traditional LLMs, SMoE is based on a collection of smaller, specialized experts to tackle different aspects of a task. It allows efficient computation while maintaining high accuracy and
- gracefully handles a context of 32k tokens.
- handles English, French, Italian, German, and Spanish.
- shows strong performance in code generation.
- can be fine-tuned into an instruction-following model that achieves a score of 8.3 on the MT-Bench.
Mixtral 8x7B: Performance
Mixtral is a decoder-only model where the feedforward block picks from a set of eight distinct groups of parameters. The Mistral AI blog says, “At every layer, for every token, a router network chooses two of these groups (the “experts”) to process the token and combine their output additively.
This technique increases the number of parameters in a model while controlling cost and latency, as the model only uses a fraction of the total set of parameters per token. Concretely, Mixtral has 46.7B total parameters but only uses 12.9B parameters per token. It, therefore, processes input and generates output at the same speed and for the same cost as a 12.9B model.”
Mistral 7B Outperforms LLaMA 2 and GPT-3.5 by running 6x faster
Mixtral is pre-trained on data extracted from the open Web and masters French, German, Spanish, Italian, and English. Mistral AI compared Mixtral with the Llama 2 family and the GPT3.5 base model. And it matches or outperforms Llama 2 70B, as well as GPT3.5, on most benchmarks.
Source: Mistral AI
In the above figure, Mistral measures the quality versus inference budget tradeoff. Mistral 7B and Mixtral 8x7B belong to a family of highly efficient models compared to Llama 2 models. It displays more positive sentiments than Llama 2 on BOLD, with similar variances within each dimension.
How do I access Mistral 8x7B?
Like any large language model, Mixtral utilizes prompts to understand requests and produce outputs. Depending on the chosen platform and programming language, you need to set up libraries and dependencies for interacting with the model. Also, you can explore different versions and choose the one that suits your needs. According to the official Mistral website, “To enable the community to run Mixtral with a fully open-source stack, we have submitted changes to the vLLM project, which integrates Megablocks CUDA kernels for efficient inference.”
Currently, Mixtral 8x7B is available in beta. Register to get early access to all generative and embedding endpoints. It can ban some outputs from constructing applications that require a strong level of moderation, as exemplified here with proper preference tuning.
In conclusion, Mixtral 8x7B, with its open-source nature, remarkable performance, and diverse capabilities, is a revolutionary innovation in the LLM landscape. It aims to explore the power of AI and push the boundaries of both users and organizations. As the Mixtral promises to evolve and modify, it can be transformative, leading to a new era of human-AI collaboration.
