Mixtral of Experts
We introduce Mixtral 8x7B, a Sparse Mixture of Experts (SMoE) language model. Mixtral has the same architecture as Mistral 7B, with the difference that each layer is composed of 8 feedforward blocks (i.e. experts). For every token, at each layer, a router network selects two experts to process the current state and combine their outputs. Even though each token only sees two experts, the selected experts can be different at each timestep. As a result, each token has access to 47B parameters, but only uses 13B active parameters during inference. Mixtral was trained with a context size of 32k tokens and it outperforms or matches Llama 2 70B and GPT-3.5 across all evaluated benchmarks. In particular, Mixtral vastly outperforms Llama 2 70B on mathematics, code generation, and multilingual benchmarks. We also provide a model fine-tuned to follow instructions, Mixtral 8x7B - Instruct, that surpasses GPT-3.5 Turbo, Claude-2.1, Gemini Pro, and Llama 2 70B - chat model on human benchmarks. Both the base and instruct models are released under the Apache 2.0 license.
PDF AbstractCode
Natural Questions
MMLU
GSM8K
TriviaQA
HumanEval
HellaSwag
MATH
PIQA
WinoGrande
The Pile
ARC (AI2 Reasoning Challenge)
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
|---|---|---|---|---|---|---|
| Common Sense Reasoning | ARC (Easy) | Mistral 7B (0-shot) | Accuracy | 80.5 | # 11 | |
| Common Sense Reasoning | ARC (Easy) | Mixtral 8x7B (0-shot) | Accuracy | 83.1 | # 9 | |
| Code Generation | HumanEval | Mistral 7B (0-shot) | Pass@1 | 26.2 | # 87 | |
| Code Generation | HumanEval | Mixtral 8x7B (0-shot) | Pass@1 | 40.2 | # 61 | |
| Math Word Problem Solving | MATH | Mixtral 8x7B (maj@4) | Accuracy | 28.4 | # 67 | |
| Math Word Problem Solving | MATH | Mistral 7B (maj@4) | Accuracy | 12.7 | # 86 | |
| Parameters (Billions) | 7 | # 58 | ||||
| Code Generation | MBPP | Mixtral 8x7B (3-shot) | Accuracy | 60.7 | # 35 | |
| Multi-task Language Understanding | MMLU | Mixtral 8x7B (5-shot) | Average (%) | 70.6 | # 22 | |
| Question Answering | PIQA | Mixtral 8x7B (0-shot) | Accuracy | 83.6 | # 9 | |
| Question Answering | PIQA | Mistral 7B (0-shot) | Accuracy | 82.2 | # 16 | |
| Common Sense Reasoning | WinoGrande | Mixtral 8x7B (0-shot) | Accuracy | 77.2 | # 17 | |
| Common Sense Reasoning | WinoGrande | Mistral 7B (0-shot) | Accuracy | 74.2 | # 25 |
