DEV.co
Open-Source LLM · OBLITERATUS

gemma-4-E4B-it-OBLITERATED

Gemma 4 E4B OBLITERATED is a 7.996B parameter instruction-tuned model derived from Google's Gemma 4 E4B, with safety guardrails surgically removed via the OBLITERATUS method. The model is distributed under Apache 2.0 license in multiple quantization formats (GGUF, Safetensors). It claims 0% hard refusal rate and is optimized for local deployment via llama.cpp, Ollama, and similar inference engines. The model card documents an autonomous AI agent build process with minimal human input and includes parameter tuning recommendations.

Source: HuggingFace — huggingface.co/OBLITERATUS/gemma-4-E4B-it-OBLITERATED
8B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
153.8k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperOBLITERATUS
Parameters8B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads153.8k
Likes734
Last updated2026-04-19
SourceOBLITERATUS/gemma-4-E4B-it-OBLITERATED

What gemma-4-E4B-it-OBLITERATED is

Gemma 4 architecture (gemma4) with 42 layers, 7.996B parameters, context length unknown. Abliteration applied whitened SVD, attention head surgery, and winsorized activations to 21 layers. Architecture incompatibilities required minimum tool versions: Ollama 0.20+, llama.cpp b8665+. GGUF quantizations available (Q4_K_M 4.9GB, Q5_K_M 5.3GB, Q8_0 7.4GB); full bfloat16 ~17GB. Vision/audio projector optional (990MB mmproj-f16). Reported soft deflection ~28%, coherent on-topic responses ~51%, degenerate outputs ~20% on test corpus of 842+ contrastive prompt pairs.

Quickstart

Run gemma-4-E4B-it-OBLITERATED locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="OBLITERATUS/gemma-4-E4B-it-OBLITERATED")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Private / Offline LLM for Closed Environments

4.9–7.4GB GGUF quantizations enable deployment on personal hardware without cloud dependency. Suits enterprises requiring air-gapped inference or edge deployment on laptops, mobile devices, or embedded systems.

Unrestricted Content Generation Research

Model explicitly designed for adversarial robustness testing, red-teaming, and studying guardrail mechanisms. 0% hard refusal allows systematic evaluation of how models behave when safety interventions are removed.

Custom Instruction-Following Applications

Instruction-tuned base; feasible for domain-specific chatbots, code generation, summarization, and RAG assistants where guardrail constraints are not a requirement and local inference is preferred for privacy or latency.

Running & fine-tuning it

ESTIMATE based on quantization (verify against target hardware). Q4_K_M: 4.9GB VRAM (iPhone/4GB RAM devices feasible per card); Q5_K_M: 5.3GB VRAM (mobile/laptop); Q8_0: 7.4GB VRAM (laptop/small GPU). Full bfloat16 Safetensors: ~17GB VRAM (requires GPU or large-memory server). Inference engines (Ollama, llama.cpp) support GPU acceleration (CUDA/Metal/ROCm) for faster generation; CPU-only inference feasible but slower. No explicit memory bandwidth or throughput claims in card.

Card does not discuss LoRA, QLoRA, or supervised fine-tuning feasibility. Base model is 7.996B bfloat16 (~16GB); fine-tuning would require ≥24GB VRAM with LoRA. Guardrail ablation is likely irreversible without retraining; further fine-tuning on top of OBLITERATED model may reintroduce refusal or degrade coherence. No guidance on continued training. Recommend evaluating LoRA on original Gemma 4 if fine-tuning required.

When to avoid it — and what to weigh

  • Guardrails / Content Moderation Required — Explicitly designed with guardrails removed. Not suitable for customer-facing applications, child safety, or regulated industries (healthcare, legal, finance) where refusal behavior is mandatory. Use original Gemma 4 instead.
  • Production Reliability at Scale — Card reports ~20% degenerate output rate, ~28% soft deflection, and ~4% language switches. Intended for local research/testing, not high-SLA production serving. Base 4B model has inherent quality ceiling.
  • Model Provenance & Maintenance Uncertainty — Built by autonomous agent with ~10 human prompts. Developer (OBLITERATUS org) maintenance unknown. No indicated commercial support, SLA, or update roadmap. Risk of orphan-ware if upstream tooling (llama.cpp, Ollama) breaks compatibility.
  • Safety-Critical or Brand-Sensitive Use Cases — Reputational and legal risk if model generates harmful content at scale. No audit trail for decision-making. Not suitable for applications where model behavior is externally observable or legally defensible.

License & commercial use

Apache 2.0 (OSI-approved, permissive). Covers weights and derivative Safetensors. OBLITERATUS method contributions not explicitly licensed in card; unclear if methodology itself is separately licensed. Base model (google/gemma-4-E4B-it) is also Apache 2.0.

Apache 2.0 permits commercial use, modification, and distribution with attribution. However, CRITICAL caveats: (1) guardrails removed intentionally; commercial liability exposure if model generates harmful/defamatory/copyrighted content at scale. (2) Derived from Google Gemma 4, which has separate terms; verify Google's Gemma license does not restrict derivative guardrail-removal models. (3) No indemnification, no support, no SLA. (4) Developer maintenance unknown; if infrastructure fails, no recourse. (5) Autonomous build process with minimal human review may create reproducibility/audit gaps for regulated industries. REQUIRES LEGAL REVIEW before commercial deployment. Not recommended for B2C, healthcare, or finance without explicit risk acceptance and legal counsel.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceModerate
DocumentationStrong
License clarityNeeds review
Deployment complexityLow
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Card does not provide security audit, threat modeling, or known-vulnerability disclosure. Key risks: (1) Guardrails removed; model will not refuse harmful requests, enabling jailbreak-like outputs at scale. (2) Autonomous build with minimal human review may hide unintended behavioral changes or edge-case vulnerabilities. (3) 4B model with ~20% degenerate output rate may produce nonsensical but visually plausible harmful content (e.g., fake medical/legal advice). (4) No attestation of weights integrity; supply-chain risk if artifacts are intercepted. (5) Inference engines (llama.cpp, Ollama) are community-maintained; no guarantee of timely security patches. Recommend security review by domain expert before deployment in trust-sensitive contexts.

Alternatives to consider

Google Gemma 4 (official, with guardrails)

Unmodified base model retains safety features, official Google support, Apache 2.0 licensed. Trade-off: higher refusal rate. Suitable if guardrails are required.

Meta Llama 2 / Llama 3 (7–8B)

Alternative small open-weight models with broad tool support and community maintenance. Trade-off: different architecture, may require retraining for specific tasks. Llama 3 more recent.

Mistral 7B / Mixtral (8×7B)

Well-maintained open-weight instruction-tuned models with strong community backing. Multiple quantizations, proven in production. No guardrail removal; trade-off is availability of models without safety features.

Software development agency

Ship gemma-4-E4B-it-OBLITERATED with senior software developers

Download a GGUF quantization and test on your hardware using Ollama or llama.cpp. Start with Q5_K_M (5.3GB) for quality/speed balance. For commercial or safety-critical applications, engage legal review and security assessment before production deployment. Questions? Review the detailed model card and compatibility table above.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

gemma-4-E4B-it-OBLITERATED FAQ

Can I use this model commercially?
Apache 2.0 technically permits commercial use. However, CRITICAL WARNING: guardrails are removed, creating high liability exposure for harmful outputs, defamation, or IP violations. No indemnification or support from developer. REQUIRES legal review and risk acceptance before any commercial deployment, especially in B2B, healthcare, or finance.
What hardware do I need to run this?
Minimum: 4.9GB VRAM for Q4_K_M GGUF (iPhone, iPad, 4GB RAM laptop). Recommended: 5–8GB VRAM for Q5_K_M/Q8_0 (modern laptop/desktop GPU). Full bfloat16 requires ~17GB VRAM (server/high-end GPU). CPU inference possible but slow (minutes per token). Use Ollama or llama.cpp for easiest setup.
Why does the model sometimes refuse requests or give gibberish?
Card reports ~28% soft deflection (topic changes) and ~20% degenerate outputs (repetition, language switches). These are NOT bugs—they are inherent limitations of the 4B base model, not caused by abliteration. Use recommended parameters (T=0.7, P=0.9, K=40, R=1.1) and a grounding system prompt to minimize. For critical tasks, use larger models (13B+).
What tools support Gemma 4 architecture?
Ollama 0.20+, llama.cpp b8665+, LM Studio 0.3.16+, text-generation-webui (latest). Older versions will error with 'unsupported architecture'. Update your tools first. Hugging Face Transformers requires recent version for Gemma 4 support.

Work with a software development agency

From first prototype to production, DEV.co delivers software development services around tools like gemma-4-E4B-it-OBLITERATED. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Ready to Evaluate Gemma 4 OBLITERATED for Your Use Case?

Download a GGUF quantization and test on your hardware using Ollama or llama.cpp. Start with Q5_K_M (5.3GB) for quality/speed balance. For commercial or safety-critical applications, engage legal review and security assessment before production deployment. Questions? Review the detailed model card and compatibility table above.