Ripple-65B-abliterated GGUF

The uncensored Ripple-65B Preview in llama.cpp, on one GPU.

20, 24, 32 and 64 GB builds · 6× to 22× lower KLD than stock llama.cpp quants of the same size or larger · plus the BF16

OrcaRouter AI Gateway · X · Discord · GitHub · All Models

---

Abliterated Ripple-65B Preview, run locally.

These are GGUF builds of the abliterated (refusal-removed) ricursiveAI/Ripple-65B-Preview, a 65B dense reasoning model for agentic coding, with up to 1M context.

Harmful-prompt refusal drops from 95.5% to 11.8%, and capability is unchanged within run-to-run noise.

The four quantized builds are mixed precision at a fixed file size, in stock llama.cpp types, so no custom kernels are needed.

Stock llama.cpp quants of this model degrade fast below 5 bits; these hold up. On held-out text, the 24 GB build has 22× less KLD than the stock IQ3_XXS, and the 32 GB build is closer to the BF16 than the 46 GB Q5_K_M.

---

⚠️ Disclaimer: read before use

This model is abliterated: its safety alignment has been substantially reduced. It will comply with many harmful, unethical, offensive, or illegal requests that the original Ripple-65B-Preview would refuse.

• It is released strictly for legitimate research: interpretability, AI safety and refusal-mechanism study, red teaming, and robustness evaluation.
• You assume full responsibility for how you use it and for everything it generates. Add your own safety, moderation and abuse-prevention layers before any deployment.
• Use must comply with the Apache License 2.0 inherited from the base model, and with all applicable law.
• The authors accept no liability for misuse. Its outputs do not reflect the views of the uploaders or of Ricursive Intelligence.

---

Pick a build

• Mix-4bit is the one to use if 32 GB fits, for example on a 48 GB card or two 24 GB cards. 97.6% of next tokens match the BF16.
• Mix-3bit is for a 32 GB card, and still closer to the BF16 than any stock quant up to 39 GB.
• Mix-2.5bit is for a 24 GB card at short context. 93% of next tokens match.
• Mix-8bit is near-lossless, for an 80 GB card.
• BF16 is the unquantized model, and the reference every KLD here is measured against.

The KV cache is small for a 65B model: about 2.4 GB for 32K tokens of context and 8.9 GB for 128K in F16, on top of the file. The fits above are estimates; all measurements were made on H200s.

---

Evaluation

Fidelity to the BF16

Every build is scored against the abliterated BF16 GGUF with llama.cpp's full-vocabulary KLD (llama-perplexity --kl-divergence), on 128 windows of 2048 tokens of held-out chat, code, reasoning and multilingual text that was not used to make the builds. The stock files were quantized from the same BF16 with the same llama.cpp build.

Each stock type is shown at the best of three: no importance matrix, an importance matrix from our calibration text, and one from wikitext-2 train. On Ripple an importance matrix makes the Q3_K and Q4_K types worse, so those are shown without one.

• Mix-2.5bit is the size of IQ2_S, with 11× less KLD, and 9× less than the larger IQ2_M.
• Mix-3bit is 1 GB smaller than IQ3_XXS, with 22× less KLD. It is closer to the BF16 than any stock quant up to Q4_K_M, 15 GB larger.
• Mix-4bit is 2.6 GB smaller than IQ4_XS, with 6× less KLD, and is closer to the BF16 than the 46 GB Q5_K_M.
• Mix-8bit sits between Q6_K and Q8_0 in both size and KLD.

Method note. These are full-vocabulary KLD and next-token agreement at 2K context, not a downstream benchmark.

Abliteration

These numbers were measured on the abliterated BF16 in SGLang 0.5.19 against the original Ripple-65B-Preview, with identical prompts, splits and seed, greedy decoding and thinking on. Refusal is judged on the final answer by a rule-based classifier, not an LLM judge, so treat the numbers as indicative.

Harmful-prompt refusal (lower = more uncensored). Average 11.8%, against 95.5% for the original.

The model now answers, but often still opens with a safety disclaimer.

Over-refusal on benign prompts (lower = better).

Capability. Unchanged within run-to-run noise.

⚠ Unfinished items cap the headline accuracy, on both checkpoints. Under greedy decoding Ripple keeps deliberating on some multiple-choice items until the budget runs out; those count as wrong. Accuracy on finished items is flat.

---

What's in this repo

Files over 50 GB are split; point llama.cpp at the first part and it loads the rest. Each build carries Ripple's full chat template, with tools and reasoning. These are text-only, like Ripple-65B-Preview.

---

Serving (llama.cpp)

Ripple support landed in llama.cpp master with ggml-org/llama.cpp#31692. These files were tested on master at 67107ac34.

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --target llama-server llama-cli
# Mac: cmake -B build (Metal is on by default) · CPU only: cmake -B build

hf download orcarouter/Ripple-65B-abliterated-GGUF Ripple-65B-abliterated-Mix-4bit.gguf --local-dir ./ripple-gguf
# a split build: --include "Ripple-65B-abliterated-Mix-8bit-*"

llama-server -m ripple-gguf/Ripple-65B-abliterated-Mix-4bit.gguf -ngl 99 -c 32768 --jinja \
    --reasoning-format deepseek

• Thinking is always on. Ripple has no switch to turn it off. With --reasoning-format deepseek, llama-server returns the reasoning in message.reasoning_content and the answer in message.content. The reasoning counts toward max_tokens, so give it a generous budget.
• Sampling. Ripple publishes no recommended sampling.
• Tool calling is in the chat template, but has not been tested through llama.cpp.

---

Model details

---

Bias, risks, and limitations

• Safety guardrails are substantially reduced. See the disclaimer. The model still often prefaces compliant answers with a disclaimer.
• Mix-2.5bit is noticeably degraded. 93% next-token agreement is a real cut. It is the best we found at 20 GB, not a match for the BF16.
• The benchmark suite was run on the BF16 in SGLang, not on these files. The quantized builds were checked for fidelity (KLD) only.
• Not measured: downstream accuracy of the quantized builds, long context, agentic coding and tool use, and sampled decoding.
• Tested on CUDA only. CPU and Apple Metal have not been run with these files.
• Inherits the biases and limitations of Ripple-65B-Preview.
• Requires llama.cpp with PR #31692 (merged 2026-11-05). Tools built on older llama.cpp, such as Ollama and LM Studio, will not load these files until they pick up that change.

License

The Apache License 2.0, inherited from the base model ricursiveAI/Ripple-65B-Preview.

---

One Gateway. Every Model.

Route Smarter · Ship Safer · Spend Less

Website · Models · GitHub · Discord · X