In plain words: Each layer holds eight expert networks, and a router picks two per word, so the model stores 47 billion parameters but uses 13 billion. It matched or beat Llama 2 70B and GPT-3.5 on every test, with much bigger leads in math and code.
Abstract · 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.
Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, et al.
arXiv:2401.04088 · cs.LG, cs.CL · submitted Jan 8, 2024
abstract · pdf · html · See more details at https://mistral.ai/news/mixtral-of-experts/
Something that has come to light since the release of the weights, and not mentioned in this paper is that it looks like fairly likely that the 8 experts were all seeded by Mistral 7B and subsequently diverged. This has generated a lot of experimentation in the local LLM community with cloning models as a way to cheaply generate experts.
It was generally thought likely that training an 8x7B network would be as much work as training 8 7B networks, but this seems not to have been true for Mistral, which is super interesting.
There's still a lot of rapid innovation happening in this space, with papers like Calm from DeepMind this week, and a lot of the adhoc experimental layer combining happening in the wild, (see, e.g. Goliath-120b), I think we're likely to see some pretty interesting architectural improvements this year in the LLM space.
Calm seems to point the way to a next step after MoE, and models like Goliath seem to indicate that even a really really lazy version of Calm (no Linear layer combination, just literally alternating layers at full weights) can be very impactful. Overall I think we will see really, really strong models that are performant on consumer hardware in 2024, likely first half of this year.