Model catalogue - larger open weight models and more powerful GPU

[ndjones]

Hi, we're building and running inference for researchers, focused on long context and tool use, on a small fleet of HGX H200, B200, RTX Pro 6000. We're looking at a mixture of model sizes including at least one at the frontier scale. Are you planning to extend your model catalogue into this scale, or if not could you create a CONTRIBUTING.md to describe how the community might provide their own? thanks!

[svonava]

Hi @ndjones I'm improving the qwen3.6 and gemma 4 profiles for h100 and h200s, with / without thinking and 256k / 32k context sizes, should land tomorrow-ish:

For larger models, is there anything in particular you have in mind?

I've been looking at https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 as the first potential candidate to venture into ~300B param area, basically better at everything vs GLM5.2 which is 753B frontier from literally 2 months ago. H200 could be capable of serving that at Q4.

[svonava]

Also @ndjones we are planning to move the whole development process from a private repo (right now we build in private and release into superlinked/sie) here into public - will happen along with CONTRIBUTING + you will see all the in-flight PRs, issues and so on.

Should be quite interesting, given this:

[ndjones]

looks good! we're benchmarking deepseek v4, glm 5-2, and minimax m3 at the larger end, otherwise Gemma 4 / Qwen 3.6. once you're setup for contributions, we'll take a look and see if we can offer anything of use. thanks @svonava !