ricursiveAI

/

Ripple-65B-Preview

Like 383

Follow

Ricursive Intelligence 255

Ripple-65B Preview now runs in the CMF format on cortiq, a Rust inference engine with no Python and no ML framework (Vulkan, DX12, Metal, CPU). The files are at https://huggingface.co/infosave/Ripple-65B-Preview-cmf (Apache-2.0 for the engine; the model's own license applies to the weights).

Files

Every file carries a header that keeps device memory in place on long prompts (the far field of our O(1) attention paper) and keeps passwords and codes in the text exact. The two-bit file keeps the busiest 5 % of each layer's FFN neurons and the K and V projections at eight bits.

Measured (cortiq bench --core, 60 tokens after the prompt)

• NVIDIA A40 (46 GB), the four-bit file, prompts of 4 000 / 16 000 / 32 000 / 64 000 tokens: 12.0 / 11.7 / 11.8 / 11.8 tok/s at 35.3 / 36.6 / 36.6 / 36.6 GB of VRAM. The same file without the header falls to 8.8 tok/s at 43.2 GB by 64 000 tokens.
• NVIDIA RTX PRO 6000 (96 GB), a 4 000-token prompt: q4tp 209 tok/s prompt, 29.8 tok/s decode; q2tp 155 / 27.8; q8_2f 156 / 18.3.

The engine maps the architecture as published: gated attention, the value residual from the first layer, sandwich norms, the clamp on the SwiGLU, 16 full and 64 sliding layers. Greedy outputs of the q4tp file were checked against the plain path token by token. A DFlash 2 drafter converted from the published checkpoint comes beside the model as an exact sidecar (CMF_DFLASH=1); its rounds do not pay yet on our cards, which the card says plainly.

Run

cargo install cortiq-cli
hf download infosave/Ripple-65B-Preview-cmf ripple-65b-q4tp.cmf --local-dir .
cortiq run ripple-65b-q4tp.cmf --prompt "Explain the Fourier transform in three paragraphs." --greedy
cortiq serve ripple-65b-q4tp.cmf --port 8080   # OpenAI-compatible API

The card has the full tables, the figures and the recipe that builds the files from the BF16 checkpoint. Corrections to the mapping are welcome.