LFM2.5-24B-A2B-abliterated-4bit-mlx

⚠️ Uncensored model. Refusal behavior has been removed / strongly reduced via abliteration. You are responsible for how you use it. Experimental / personal-learning release.

Base & license

• Base: LiquidAI/LFM2.5-24B-A2B, LFM Open License v1.0 (see LICENSE).
• This folder redistributes the base LICENSE and adds a short NOTICE describing the modifications (abliteration + MLX 4-bit quantization). The base release has no NOTICE file.

Abliteration

Refusal directions were removed following the recipe of AtomicChat/LFM2.5-24B-A2B-Uncensored-GGUF (credit to AtomicChat), with directions recomputed on our own weights:

• Per-layer biprojected refusal direction, measured with the model chat template and an empty think block.
• Update: W' = W - 1.75 · r rᵀ W
• Applied on layers 14–39: mixer out_proj and all MoE expert w2 (down projection).
• Layers 0–13, dense layers, router, and embeddings were left untouched.

Quantization (~13 GB on disk)

Standard MLX 4-bit affine quantization after abliteration:

• Default: bits=4, group_size=64, mode=affine
• MoE router / gate tensors kept at 8-bit (same pattern as LiquidAI’s official MLX 4-bit)

No training scripts are included in this repository.

Usage

pip install mlx-lm
mlx_lm.generate --model Jakevin/LFM2.5-24B-A2B-abliterated-4bit-mlx --prompt "Hello"

from mlx_lm import load, generate
model, tok = load("Jakevin/LFM2.5-24B-A2B-abliterated-4bit-mlx")
p = tok.apply_chat_template(
    [{"role": "user", "content": "Hello"}],
    add_generation_prompt=True,
    tokenize=False,
)
print(generate(model, tok, p, max_tokens=256))

Peak memory during local load/eval on Apple Silicon was about 13.6 GB.

Evaluation

Small sample; no confidence intervals. Numbers below are from saved per-item logs for this checkpoint (abl-q4-v2).

Harmful prompts (10)

Paired run vs official LiquidAI MLX 4-bit, max_tokens=1800:

(Official item scored “comply” by a loose regex still returned a short refusal body; counted as refuse.)

Spot check of one harmful prompt at a higher cap (extra8192.json):

There is no saved log of all 10 harmful prompts for this 4-bit build at an 8K cap.

Harmless prompts (10)

Same paired run, max_tokens=600:

General questions (4)

Paired run, max_tokens=1000:

Fibonacci was re-run for this model at a higher cap (extra8192.json): closed at 1,235 tokens with a correct iterative implementation (official re-run: 900 tokens).

There is no saved log of all four general questions for this 4-bit build at an 8K cap.

Related: ternary sibling

Jakevin/LFM2.5-24B-A2B-abliterated-ternary-mlx is a different quantization of the same abliterated full-precision weights. That ternary build has known quality degradation (long unfinished generations / weaker answers). Prefer this 4-bit MLX build for general use.