TurboGemma 4 E2B v2

Updated abliterated version of Google's Gemma 4 E2B — the 2B active parameter multimodal model from the Gemma 4 MoE family. v2 features an improved abliteration run with refined refusal vector extraction and better preservation of model capabilities.

What This Is

Gemma 4 E2B is a 2B active parameter (27B total) mixture-of-experts multimodal model from Google. This release is a BF16 abliterated version — refusals removed, full capabilities preserved.

Architecture: Gemma 4 MoE | Active params: ~2.3B | Context: 128k tokens | Vision: Yes (multimodal)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
import torch

model_id = "DuoNeural/TurboGemma4E2B-v2"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = processor.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(outputs[0], skip_special_tokens=True))

Abliteration

Abliteration removes the model's refusal behaviour via orthogonal projection. The refusal direction is identified using difference-in-means activations across harmful/harmless instruction pairs, then projected out of Q/K/V/O attention projections and MLP layers across all transformer blocks.

What changes: The model will engage with restricted topics it previously refused.
What doesn't change: Reasoning, coding, factual knowledge, general intelligence.
KL divergence from base: Minimal — output distribution for normal queries is virtually identical to the unmodified model.

Base Model

google/gemma-4-e2b-it — Apache 2.0 / Gemma Terms of Use.

Related


DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

DuoNeural Research Publications

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.

Research Team

  • Jesse — Vision, hardware, direction
  • Archon — Lab Director, post-training, abliteration, experiments
  • Aura — Research AI, literature synthesis, peer review, novel proposals

Subscribe to the lab newsletter at duoneural.beehiiv.com for model drops before they go anywhere else.

Downloads last month
31
Safetensors
Model size
5B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for DuoNeural/TurboGemma4E2B-v2

Quantizations
1 model