Ripple-65B-Preview — Q8_0
Q8_0 GGUF quantization of Ricursive's original BF16 model. The source has 64,782,689,360 total parameters. Weight files total 64.11 GiB; inference needs additional memory.
Runtime
Requires the experimental Ricursive implementation from llama.cpp PR #31692, tested at commit 6414d5b0c9e5895473036c90c2bd305ce4ceccc4. This port is unmerged at the time of this release; compatibility with stable llama.cpp, Ollama or LM Studio has not been established.
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git fetch origin pull/31692/head
git checkout 6414d5b0c9e5895473036c90c2bd305ce4ceccc4
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j 6 --target llama-completion
hf download Chungulus/Ripple-65B-Preview-Q8_0-GGUF --local-dir ./ripple-Q8_0
./build/bin/llama-completion -m ./ripple-Q8_0/Ripple-65B-Preview-Q8_0-00001-of-00002.gguf -cnv -st --jinja -c 2048 -ngl 0 --no-repack --load-mode mmap -p 'Hello!'
For split files, download every shard and load the first shard shown above. This Q8_0 artifact was sanity tested with CPU inference, memory mapping and weight repacking disabled. The command avoids a full anonymous-memory repack copy. Choose offload settings for available memory.
Provenance and checks
Each format is quantized independently from a BF16 GGUF converted from the original source revision e27e171792f303d74c048184e8259ba663cfa0d5. No requantization, calibration dataset or importance matrix was used. The source chat template and end-of-turn token are preserved.
The converter/runtime matched the unmodified vendor forward on a random ten-layer fixture and a five-layer crop of the original weights, including cached decoding. Every converted source tensor's shape and up to four scalar samples matched. This does not measure full-model benchmark quality.
This artifact passed short deterministic arithmetic, factual-answer and JSON-format generation checks. See validation.json for actual outputs, methods and limits. Upstream benchmark scores are not reported as measured scores for this quantization. quantizatia-manifest.json records source/tool revisions, original-shard digests and release-file SHA256 checksums.
Apache 2.0; original model and license belong to Ricursive Intelligence and their respective authors. This repository distributes a community quantization, not an official Ricursive release.