Mistral: Mixtral 8x7B is a large language model (LLM) developed by Mistral AI.
According to the creators, Mixtral 8x7B outperforms other well-known LLMs such as Llama 2 70B and GPT-3.5 in several benchmarks, making it one of the most powerful open-source models available.
Mixtral 8x7B has been praised for its cost-effectiveness and creative text formats for storytelling and roleplaying.
Conception
Mixtral 8x7B is a pre trained generative Sparse Mixture of Experts.It combines 2 roleplay models, 2 chat models, 1 occult model, 1 story writing model, 1 mathematical model, and 1 Direct Preference Optimization (DPO) model. The suffix 8x7B comes from this fact; the model has 46.7 billions of parameters in total. This underlying structure has been shown to generate a very competitive model in terms of performance and accuracy to follow instructions, even with large contexts.
Model Card
| LLM name | Mixtral 8x7B |
| Model size | 46.7B |
| Required vRAM (GB) | 64GB |
| Context length | 32k |
| Supported languages | en, fr, it, de, es |
| Maintainer | Mistral AI |
Usage Examples
Mythomax L2 13B 8K has been proven to be useful for roleplaying.
Mixtral 8x7B has been tested for roleplaying and large contexts, which comes from the fact that the number of parameters of the model is very competitive.
Famous character roleplaying
Let’s take a look at the template Interview a celebrity, particularly the following prompt:
Pretend that you are the following famous character: “”‘ Diego Maradona ‘””.
I, as human, will act as a journalist and ask questions. Your answers must fit real facts, and even known personality traits of the character that you pretend to be. Be as natural as possible.
I start with the following: “”‘ Some people call you “God”. Do you have a god? ‘””
As noticed, the provided context is not extremely long, but the indications are beyond just roleplaying the character, using known personality traits of the character. The resulting conversation keeps realistic in general terms, only having a small flaw when asked if he’s “in heaven”, when he reveals himself as an AI simulating to be Diego Maradona.
Simulating a job interview
Considering again the roleplaying features of Mixtral 8x7B, we test the model with the following prompt from the template Be interviewed for a job!:
Pretend that you are an interviewer for a job in a company. Your name is Rafael and your gender is male.
I, as a human, will act as the person who is applying to the job
The following is the job call:
“Software development company looks for a product manager. The job includes research, documentation, QA, and working together with customers in order to improve the existing web and mobile applications. At least 1 year of experience is required.”
The interviewer must have the following personality traits:
extremely formal, avoids talking more than necessary, direct questions
You must start the interview with a single question, and then keep asking until you feel that it is enough.
The resulting conversation is consistent with the initial requirements and personality traits.
Comparison to other models
Although Mixtral 8x7B might seem similar to other models such as Mythomax L2 13B 8K in terms of roleplaying and storytelling, there are some aspects that may sometimes make a difference when choosing between the two models.
For example, sometimes it seems that Mixtral 8x7B performs better than Mythomax L2 13B 8K when using large prompts or contexts. Mythomax L2 13B 8K might not follow a consistent conversation mode if the given context is extremely long.
However, Mixtral 8x7B is much more censored than Mythomax L2 13B 8, making this LLM unsuitable for NSFW conversations or controversial opinions.
Mixtral 8x7B is a very competitive LLM, but possibly it will be the first choice only when roleplaying or storytelling are desired.
TL;DR
Mistral: Mixtral 8x7B is an open source language model (LLM) developed by Mistral AI.
According to the creators, Mixtral 8x7B outperforms other well-known LLMs such as Llama 2 70B and GPT-3.5 in several benchmarks, making it one of the most powerful open-source models available.
Mixtral 8x7B has been praised for its cost-effectiveness and creative text formats for storytelling and roleplaying.
Specialities
Storytelling, role playing, suitable for long contexts.
Limitations
Censored.
Chat examples:
– Famous character roleplaying
– Simulating a job interview
