•

Mistral AI just dropped their new model, Mistral Large 2

Large Enough

https://mistral.ai/news/mistral-large-2407/

Today, we are announcing Mistral Large 2, the new generation of our flagship model. Compared to its predecessor, Mistral Large 2 is significantly more capable in code generation, mathematics, and reasoning. It also provides a much stronger multilingual support, and advanced function calling capabilities.

Another day, another model.

Just one day after Meta released their new frontier models, Mistral AI surprised us with a new model, Mistral Large 2.

It's quite a big one with 123B parameters, so I'm not sure if I would be able to run it at all. However, based on their numbers, it seems to come close to GPT-4o. They claim to be on par with GPT-4o, Claude 3 Opus, and the fresh Llama 3 405B regarding coding related tasks.

benchmarks

It's multilingual, and from what they said in their blog post, it was trained on a large coding data set as well covering 80+ programming languages. They also claim that it is "trained to acknowledge when it cannot find solutions or does not have sufficient information to provide a confident answer"

On the licensing side, it's free for research and non-commercial applications, but you have to pay them for commercial use.

4 comments

StubbornCassette8

•

What are the hardware requirements on these larger LLMs? Is it worth quantizing them for lower-end hardware for self hosting? Not sure how doing so would impact their usefulness.

hohoho

•

The general rule of thumb that I’ve heard is that you need 1GB of memory for ever 1B parameters. In practice however I’ve found this to not be the case. For instance on a GH200 system I’m able to run Llama3 70b in about 50GB of memory. Llama3.1 405b on the other hand uses +90GB of GPU memory and spills over to using about another 100GB of system memory… but runs like a dog at 2 tokens per second. I expect inference costs will come down over time but for now would recommend Lambda Labs if you don’t have the need for a GPU workstation.

bazsalanszky

•

From what I've seen, it's definitely worth quantizing. I've used llama 3 8B (fp16) and llama 3 70B (q2_XS). The 70B version was way better, even with this quantization and it fits perfectly in 24 GB of VRAM. There's also this comparison showing the quantization option and their benchmark scores:

1000029570

Source

To run this particular model though, you would need about 45GB of RAM just for the q2_K quant according to Ollama. I think I could run this with my GPU and offload the rest of the layers to the CPU, but the performance wouldn't be that great(e.g. less than 1 t/s).

anticurrent

•

Wake me up when I can ran that locally on my potato Laptop.