Welcome to the Runpod Roundup for the week ending July 29, 2023. In this issue, we'll be discussing the newest advancements in AI models over the past week, with a focus on new offerings that you can run in a RunPod instance right this second. In this issue, we'll be looking at the new SDXL release as well as new LLM model advancements.
High Context Llama-2 Models Now Available
It's not even been a week since Meta and Microsoft released Llama-2, and the community has been hard at work...
togethercomputer has released their 32k token context version of Llama-2 7b for anyone to download off of Huggingface. Although there have been several versions of closed models (GPT-4, et al) that have had 32k token context, this is I believe the first freely available open-sourced model that has made the jump to 32k. Although many front ends such as Oobabooga do not yet support 32k context windows, that is likely to change as these models become more commonplace.
(Could OpenAI also be looking into releasing an open-source LLM of their own in the not-too-distant future? Maybe...)
Stable Diffusion XL 1.0 Released For All
The long-awaited SDXL is finally out and available for use on RunPod - no research credentials required! Not to belabor the point, but SDXL really is the next generation (pun intended) of AI art, and we've got a helpful blog post that will help you get it up and running in your pod.
StabilityAI Releases Stable Beluga 1 and 2 LLMs
It's clearly been a busy week for Stability AI, as they have also released two LLMs, Stable Beluga 1 and 2. The Stable Beluga 1 entry on Huggingface only holds delta weights (likely due to licensing issues with Llama-1) but Stable Beluga 2 is the complete model available to download. The former is a Llama-1 65b model, while the latter is a Llama-2 70b version, both of which have been finetuned on Microsoft Orca style datasets. According to their press release, both models have a focus on intricate reasoning and answering complex questions relating to specialized domains requiring subject matter expertise, such as law and mathematical problem solving.
Feel free to reach out to RunPod directly if you have any questions about these latest developments, and we'll see what we can do for you!