[feature] Support Ripple-65B-Preview (Dense Model) as a Splash package #689

Summary

Please consider adding Ripple-65B-Preview (ricursiveAI/Ripple-65B-Preview) as a supported target. Unlike most model requests, its DFlash 2 draft already exists: ricursiveAI publishes ricursiveAI/Ripple-65B-Preview-DFlash2 themselves.

It is a 65B dense model, about 40 GB at Q4_K_M and 69 GB at Q8_0, so it fits a 64 GB Mac at 4-bit and a 96–128 GB Mac at 8-bit. As a dense model that reads all of its weights for every token, it is exactly where DFlash 2 should help most on Apple silicon.

Why this model

• The draft is released by the model's authors, so no draft training is needed.
• Its model card reports leading agentic coding scores among US open-weight models (80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard), against models of 501B–975B parameters, which would make it a strong fit for Splash's coding-agent use at a size a single Mac can hold.
• It would be Splash's first large dense non-Qwen target. Its mix of 1024-token sliding-window and full-attention layers (four to one) broadens the engine beyond GDN and may overlap with the Gemma 4 work in [feature] Support Gemma 4 26B-A4B (MoE) as a Splash package #8.
• A 65B dense model decodes slowly on Macs without speculation, so DFlash 2 makes the biggest difference here.
• llama.cpp master already supports the architecture (LLM_ARCH_RICURSIVE), which gives a reference implementation to check against.

Reference

• Target: https://huggingface.co/ricursiveAI/Ripple-65B-Preview
• Draft: https://huggingface.co/ricursiveAI/Ripple-65B-Preview-DFlash2