All stories
Ripple-65B Preview Ecosystem Community reports

Ripple gets a 38.8 GB GGUF route—but stock llama.cpp cannot load it yet

AtomicChat publishes local-serving conversions and build instructions for Ricursive’s new model. Architecture support is still an open pull request.

By AITrending2 min read
Sources (2)

Why now AtomicChat created its GGUF repository November 4 at 21:11:26 UTC; llama.cpp PR31692 followed at 22:41:43 UTC and remained open at this review. This is a serving-format follow-up to the model launch, not a second model release.

Storyline / Report 2 of 2

Ripple-65B Preview launch and local serving support

Latest report
Full timeline 2 reports
  1. Community reports
    65B against the giants: Ricursive releases Ripple with an 80.52 coding claim
  2. Community reports
    Ripple gets a 38.8 GB GGUF route—but stock llama.cpp cannot load it yet (current report)

AtomicChat has published GGUF conversions of Ripple-65B Preview, including a Q4_K_M file its English card lists at 38.8 GB on disk. The practical catch is architecture support: the project says stock llama.cpp, LM Studio and Ollama cannot load these files yet.

That makes this a local-serving follow-up to Ricursive’s model launch, rather than another new model. AtomicChat’s repository was created on November 4, and the linked llama.cpp architecture proposal, PR #31692, was opened later that evening. At review it remained open and unmerged.

A build route for early adopters

The card gives two routes: build llama.cpp from the pull-request branch, or apply the repository’s ricursive-port.patch to a specific documented llama.cpp revision. Its example enables CUDA; the macOS note says to omit that flag because Metal is enabled by default. The serving example includes —jinja to use the embedded chat template with tools and reasoning.

Alongside Q4_K_M, the card lists a two-part Q8_0 conversion at 68.8 GB and a three-part BF16 conversion at 129.6 GB. Those are listed file sizes, not complete RAM or VRAM requirements. Context, cache and runtime overhead also need memory. The repository API additionally lists newer importance-matrix quantizations, but the reviewed card does not yet provide their sizes or full quality measurements.

The port author reports agreement with reference code on a small random model and the first ten real layers, plus architecture and template checks. Those are author-reported checks, not an independent evaluation of full-model coding performance. The earlier Ripple benchmark claim should not be assumed to transfer unchanged to every quantization.

The immediate recommendation is conditional: readers comfortable building an experimental serving engine now have an inspectable route and downloadable files. Readers relying on an unmodified desktop model runner should wait for explicit architecture support. The model retains the base release’s Apache-2.0 license.

Related reading