This pushed me to try pi.dev with a remote ollama instance running gemma4:26b on a 24GB RTX 3090. It is not perfect, but it's quite okay-ish with 128k context length. It's not a model primarily for coding I guess, but IIWIW. Thanks!