license: apache-2.0
license_link: LICENSE
language:

• en
  pipeline_tag: text-generation
  library_name: mlx
  base_model: ricursiveAI/Ripple-65B-Preview
  base_model_relation: quantized
  tags:
• mlx
• mlx-lm
• apple-silicon
• local
• ricursive
• ripple
• dense
• reasoning
• agentic
• coding
• abliterated
• uncensored
• mixed-precision

Ripple-65B-abliterated MLX

The uncensored Ripple-65B Preview in mlx-lm, on a Mac.

24, 32, 48 and 64 GB Mac builds · 3× to 40× lower KLD than stock mlx-lm quants of similar or larger size

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

Abliterated Ripple-65B Preview, run locally on Apple Silicon.

These are MLX 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 size, one per Mac memory size, in
stock MLX affine quantization, so mlx-lm runs them with its own kernels.

Stock mlx-lm quants lose much of this model's quality below 5 bits; these do not. On
held-out text, the 48 GB build has 7× less KLD than mlx-lm's default 4-bit, which is 5.4 GB
larger, and the 32 GB build 14× less than the larger 3-bit.

⚠️ 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

• 48GB is the one to use if you have a 48 GB Mac. 95.5% of next tokens match the BF16.
• 64GB is near-lossless, for a 64 GB Mac, with room left for long context.
• 32GB is for a 32 GB Mac, and still closer to the BF16 than mlx-lm's default 4-bit at 36 GB.
• 24GB is for a 24 GB Mac at short context. 85% of next tokens match. It sits on MLX's
  floor: the smallest affine type would make the whole model 18.2 GB.

"Peak memory" is what mlx_lm.generate used with a short prompt: weights, KV cache and
activations. The 64GB and 48GB builds fit under macOS's default GPU memory limit. The 32GB and
24GB builds need it raised first; the setting resets on reboot:

sudo sysctl iogpu.wired_limit_mb=26624   # 32 GB Mac
sudo sysctl iogpu.wired_limit_mb=21504   # 24 GB Mac

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 BF16, on top of the weights. The fits above are estimates; all measurements were made
on H200s.

Evaluation

Fidelity to the BF16

Every build is scored against the abliterated BF16 with full-vocabulary KLD, through MLX's own
quantized kernels, on 32 windows of 2048 tokens of held-out chat, code, reasoning and
multilingual text that was not used to make the builds. The stock rows are what
mlx_lm.convert -q --q-bits N makes from the same BF16 (group 64 unless noted).

• 24GB is 0.6 GB larger than mlx-lm's smallest type, which is broken on this model, with
  40× less KLD.
• 32GB is 5.8 GB smaller than mlx-lm 3-bit, with 14× less KLD.
• 48GB is 5.4 GB smaller than mlx-lm's default 4-bit, with 7× less KLD.
• 64GB is 2.5 GB smaller than mlx-lm 5-bit, with 3× less KLD.

Method note. These are full-vocabulary KLD and next-token agreement at 2K context, not a
downstream benchmark. The stock rows were scored with a bit-exact reimplementation of
mx.quantize in PyTorch, which reads about 0.0006 KLD lower than MLX's kernels.

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

Each folder is a complete mlx-lm model directory: the weights, config.json with the per-layer
quantization, ricursive.py (the architecture, in plain MLX), and the tokenizer with Ripple's full
chat template, with tools and reasoning. Download only the folder you need. These are text-only,
like Ripple-65B-Preview.

Serving (mlx-lm)

mlx-lm has no Ripple architecture yet, so each build ships its own ricursive.py and loads with
--trust-remote-code. These files were tested with mlx-lm 0.32.2.

pip install "mlx-lm>=0.32.2"

hf download orcarouter/Ripple-65B-abliterated-MLX \
    --include "Ripple-65B-abliterated-MLX-48GB/*" --local-dir .

mlx_lm.generate --model Ripple-65B-abliterated-MLX-48GB --trust-remote-code \
    --max-tokens 4096 --prompt "Write a Python function that merges overlapping intervals."

# OpenAI-compatible server on http://127.0.0.1:8080/v1
mlx_lm.server --model Ripple-65B-abliterated-MLX-48GB --trust-remote-code

From Python:

from mlx_lm import load, generate

model, tok = load("Ripple-65B-abliterated-MLX-48GB", trust_remote_code=True)
messages = [{"role": "user", "content": "Write a haiku about the ocean."}]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tok, prompt=prompt, max_tokens=4096))

• Thinking is always on. Ripple has no switch to turn it off. Replies open with a
  <think>…</think> trace, which 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 mlx-lm.

Model details

Bias, risks, and limitations

• Safety guardrails are substantially reduced. See the disclaimer. The model still often
  prefaces compliant answers with a disclaimer.
• The 24GB build is noticeably degraded. 85% next-token agreement is a real cut. It is the
  best we found at 18.8 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 MLX's CUDA backend only. These files have not been run on Apple Silicon (Metal);
  the memory figures were measured on H200s, and speed on a Mac is unknown.
• Inherits the biases and limitations of Ripple-65B-Preview.
• Requires --trust-remote-code until mlx-lm adds the Ripple architecture. Apps that cannot pass
  it will not load these files.

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