Maximizing cost effectiveness by including as many efficient models as possible

[TomLucidor]

Some examples:

• InclusionAI (Ant Group) Ring series (SWA/Linear) https://huggingface.co/inclusionAI
• Newest LFM (liquid foundation models) https://huggingface.co/LiquidAI
• nVidia's Nemotron-H (hybrid attention) family https://huggingface.co/nvidia
• Jet-Nemtron (PostNAS models) https://huggingface.co/jet-ai
• Meituan's LongCat-Flash family (Ngram-scaled) https://huggingface.co/meituan-longcat
• IBM's Granite-4.0-H (hybrid attention) https://huggingface.co/ibm-granite
• Kimi-Linear, Qwen3-Coder-Next, whatever other major firm's linear attention model
• GPT-OSS-20B, GLM-4.7-Flash, Apriel-1.5, other Qwen/DeepSeek, Gemma series, Mistral's models etc.

[1bcMax]

Thanks for putting this together — solid research.

We can absolutely add more models. Rather than overhauling the whole catalog at once, we'd rather add things that are actually useful to real workflows.

Which ones do you personally need most, and what are you using them for? Knowing the use case (long context, cheap code gen, fast simple tasks, reasoning, etc.) helps us pick the right ones and place them in the right routing tiers.

Happy to start with 2–3 models if you tell us what you'd reach for first.