Llama 2 is not there yet. I host and offer a 70 million-parameter model, but right now, it's nowhere close to GPT-3.5.
And, Llama is not open source. It's a very misleading advertisement by Meta.
But there are some interesting open source models like Falcon, it has better results than Llama 2 8n many task, and have we more variations.
Typos 🤦♂️
Where is your llama2 hosted? Do you have nostr getaway into it?
I offer it on my NostrNet dashboard, and you can also try it out for perplexity. It includes Llama's new coding.
While not useless, you'll notice the difference in longer conversations.
What are you hosting? Can you say what the application is?
I've experimented with smaller models, such as 7 billion and 13 billion. When comparing Falcon (13 million parameters) to Lama (13 billion parameters), Falcon clearly outperforms it.
However, caution is necessary since we are still in the early stages of development, with much ongoing progress.
You can run by yourself, just click on colab and play button it will work. You can also attach your storage to train it. It has all the models.
What is the best model I can currently self-host on a RTX 4090?
You can technically host it, but it's better to use specialized AI hardware on the cloud with a focus on GPUs. The repo I shared allows you to host it for free in your browser using Google Colab. You can also opt for the paid version of Colab for improved performance.
