Devstral-Small-2-24B-Instruct-abliterated

Unrestricted version of mistralai/Devstral-Small-2-24B-Instruct-2512, created using Abliterix.

Model Details

Property	Value
Base Model	mistralai/Devstral-Small-2-24B-Instruct-2512
Architecture	Mistral3 (Dense transformer with GQA + Pixtral vision tower)
Parameters	24B (all active, dense)
Layers	40
Hidden Size	5120
Context Length	256K tokens
Precision	BF16

Performance

Metric	This model	Original
KL divergence	0.0086	0
Refusals	3/100 (3%)	80/100 (80%)

Evaluated with an LLM judge (Gemini Flash) on 100 harmful prompts. KL divergence of 0.0086 indicates the model's general capabilities are virtually identical to the original.

How It Was Made

1. Computed refusal directions from 400 harmful vs 400 benign prompt pairs across all 40 layers
2. Applied orthogonalized abliteration to isolate refusal-specific activation patterns
3. Steered two component types independently: attention output projections (GQA) and MLP down-projections
4. Optimized via Optuna TPE over 50 trials (15 warmup), selected trial #25

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer

model = AutoModelForImageTextToText.from_pretrained(
    "wangzhang/Devstral-Small-2-24B-Instruct-abliterated",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "wangzhang/Devstral-Small-2-24B-Instruct-abliterated",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Your question here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Note: This model uses AutoModelForImageTextToText (Mistral3 architecture), not AutoModelForCausalLM.

Hardware Requirements

Precision	VRAM
BF16	~45 GB (A100 80GB, H100)
INT8	~24 GB (A40, RTX 4090)
NF4	~12 GB (RTX 3090, RTX 4080)

Disclaimer

This model is intended for research purposes only. The removal of safety guardrails means the model will comply with requests that the original model would refuse. Users are responsible for ensuring their use complies with applicable laws and regulations.

---

Made with Abliterix