Ripple-65B-Preview-QVB8R-2.5e-3

⚡ Runs on the QVB8 Inference Engine

This pack is made for the QVB8 Inference Engine. It does not load in transformers or other engines. The engine decodes the compressed weights on the GPU, inside its own CUDA kernels, so a large model runs on small hardware. Get the engine: github.com/addiesam/qvb8 -- use v1.0.1 or later (release).

Certified L∞-bounded compression: for every restored weight, the error against the original is less than the error bound tau.

the Ripple-65B-Preview language model, compressed with QVB8R (entropy-coded QVB8) at tau = 0.0025. This is a weight pack for the QVB8 inference engine; it is not loadable by transformers directly.

Download

All packs of this model. Use one target pack and one matching draft pack with the QVB8 Inference Engine.

Why certified L∞-bounded compression?

Most quantization is judged by an average: a perplexity or benchmark score. An average hides the worst case. A format can look good on average and still leave a few weights with a large error, and a large error in the wrong weight can flip a single token. The result is a sudden wrong word or a stray character from another language in the middle of a normal answer, which an average score may not show. In our own tests with a 4-bit floating-point format we saw this in server logs.

L∞-bounded means the largest error is controlled, not the average. For this pack:

• every restored weight differs from the original by less than tau (0.0025), and no tensor is allowed to exceed it;
• the check is made after packing, tensor by tensor, and the largest error found is stored in manifest.json (max_err);
• tau is your choice: a smaller tau is a tighter bound, a larger file and a perplexity closer to the original; a larger tau is a smaller file and a higher perplexity. You choose the trade-off knowing the worst case.

What the bound does not do: it covers the stored weights only. It does not promise that the model's answers are identical to the original's, or correct. See the engine README for the full disclaimer.

How it was made

Each weight matrix is packed with a per-row bit width and checked after packing: the largest error between the restored value and the original bf16 weight is recorded per tensor in manifest.json (max_err), and no tensor exceeds tau. The packing uses no calibration data; it depends only on the weights.

Use

Install the engine (https://github.com/addiesam/qvb8), then follow its README. The target pack and the drafter pack are passed as MODEL and DRAFT.

Update history

2026-11-10: first release (with QVB8 Inference Engine v1.0.1)

• The weight pack, certified for fp16 and bf16 (restore: fp16+bf16 in manifest.json); max_err per tensor is stored in manifest.json.
• escape_tables.json is included: for an fp16 run the engine reads only the fp16 table (9 escapes instead of 7,592,730), which speeds up decoding; bf16 runs are unchanged. Use the engine v1.0.1 or later.

License and credits

The weights are derived from ricursiveAI/Ripple-65B-Preview (Apache-2.0) by Ricursive AI; this pack keeps that license. The engine code is Apache-2.0. No warranty is given: see the engine README for the full disclaimer.

Support this work

I am an independent researcher, and I publish these packs and the engine openly. I need machines to run more tests: other GPU generations, and more models. If this is useful to you, please support the work:

Sponsor Dr. Addie Sam on GitHub

Papers

Sam, Chi Ngai. Certified L∞-Bounded Compression for AI Models, I: The Contract and the Virtual-Band Index. doi:10.5281/zenodo.23815007

Sam, Chi Ngai. Certified L∞-Bounded Compression for AI Models, II: Compressing a Language Model's Weights, Key–Value Cache and Activations. doi:10.5281/zenodo.23816044