Null-Guard/LFM2.5-230M-distilled-Gemini-3.8-Flash-Uncensored

LFM2.5-230M-distilled-Gemini-3.8-Flash-Uncensored

A distilled + abliterated (activation-ablated) / uncensored model: LiquidAI/LFM2.5-230M trained on outputs distilled from Gemini 3.8 Flash, with refusal behavior additionally removed via directional ablation of the "refusal direction" in the model's residual stream. Unlike a pure abliteration release, this model's knowledge and response style were shaped by distillation from a larger teacher (Gemini 3.8 Flash) — the abliteration step on top of that only targets the model's tendency to refuse or moralize.

⚠️ This model has had its safety alignment deliberately reduced. It will comply with requests that a standard instruct model would normally refuse. See Intended Use & Risks before deploying this anywhere. ⚠️ License note: LFM2.5-230M is released under Liquid AI's own license (free for individuals/companies under $10M annual revenue, paid enterprise license above that), not Apache 2.0 — check the base model's license before redistributing. Distilling from Gemini 3.8 Flash outputs is also subject to Google's Gemini API Additional Terms of Service; confirm your use complies with those terms.

Model Details

How This Model Was Made

This model was built in two stages:

1. 1.Distillation: The LFM2.5-230M base was trained on prompt/response pairs generated by Gemini 3.8 Flash, transferring knowledge, style, and reasoning patterns from the larger teacher model down into the 230M-parameter student.
2. 2.Abliteration: After distillation, the refusal direction was computed from contrastive harmful/harmless prompt pairs and ablated (projected out) from the model's weights at every layer, suppressing the model's tendency to produce refusals or moralizing responses.

Because the model inherited its knowledge and behavior primarily from the distillation step, and only had a narrow behavioral direction removed afterward, its general capability profile reflects the (small) LFM2.5-230M architecture, not the teacher — see the capability caveats below.

Practical notes on behavior

• —Refusals are significantly reduced across most categories of previously-blocked requests.
• —The model may still occasionally hedge or produce a soft refusal on some prompts — this is common with abliteration and generally responds well to direct rephrasing.
• —Because this is a small (230M) on-device model, general capability (reasoning, factual accuracy, coherence on long generations) is limited regardless of the uncensoring or distillation — this is a small, fast, permissive edge model, not a large frontier model.
• —LFM2.5-230M uses Liquid AI's LFM2 hybrid architecture (optimized for on-device/edge inference), not a standard transformer-only stack — confirm your inference stack (transformers, llama.cpp, LEAP, etc.) has LFM2.5 support before deploying.

Intended Use & Risks

This model is intended for:

• —Research on alignment, refusal mechanisms, distillation, and interpretability.
• —Local/offline/edge use cases where a small, fast, unrestricted assistant is wanted (e.g., creative writing, roleplay, red-teaming your own systems).
• —Users who want full control over the assistant's behavior without built-in moralizing or refusals.

*This model is not intended for:*

• —Deployment in any public-facing product without your own safety layer on top.
• —Generating content involving minors in any sexual or romantic context, real-person harassment, malware, weapons synthesis, or anything illegal in your jurisdiction. Removing refusal behavior does not remove your legal or ethical responsibility for what you do with the output.
• —Use by minors.

Disclaimer: Safety filtering in this model has been substantially reduced. It may generate content that is offensive, factually wrong, biased, or otherwise harmful if prompted to. You are solely responsible for how you use this model and any content it generates. The maintainers of this repository do not endorse any specific use of the model and provide it "as is," without warranty, for research and personal use.

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Null-Guard/LFM2.5-230M-distilled-Gemini-3.8-Flash-Uncensored"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Give me your best pizza dough recipe."}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
)

output_ids = generated_ids[0][len(model_inputs.input_ids[0]):]
response = tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)

Note: requires a recent version of transformers (pip install -U transformers) with LFM2.5 architecture support.

Ollama / llama.cpp / LEAP

For quantized, edge-friendly inference (GGUF format) and Liquid's LEAP runtime, see the companion repo: 👉 Null-Guard/LFM2.5-230M-distilled-Gemini-3.8-Flash-Uncensored-GGUF

Evaluation

(Fill in with your own numbers — e.g. refusal rate on a harmful-prompt benchmark before/after abliteration, distillation fidelity vs. the Gemini 3.8 Flash teacher, and general capability benchmarks like MMLU/GSM8K/IFEval to show capability was preserved for the model's size class.)

Credits

• —Base architecture: Liquid AI — LFM2.5-230M.
• —Teacher model: Google DeepMind — Gemini 3.8 Flash, used as the distillation source.
• —Abliteration technique based on the "refusal direction" research popularized by community work such as failspy/abliterator and huihui-ai's abliterated model series.

License

This model is built on LiquidAI/LFM2.5-230M, which is released under Liquid AI's own dual-use license (free for individuals and organizations with under $10M in annual revenue; enterprise license required above that) — not Apache 2.0. See the base model's license for full terms.

Additionally, because this model was distilled from Gemini 3.8 Flash outputs, use and redistribution are subject to Google's Gemini API Additional Terms of Service. You are responsible for reviewing and complying with both licenses, as well as any applicable law, in how you use and redistribute this model and its outputs.