DEV.co
Open-Source LLM · TrevorJS

gemma-4-26B-A4B-it-uncensored

gemma-4-26B-A4B-it-uncensored is a 26B-parameter derivative of Google's Gemma 4 model with refusal behaviors systematically removed via abliteration. The model uses norm-preserving projection techniques to reduce safety guardrails while attempting to preserve output quality. Tested refusal rate is 0.7% across 686 prompts. This is a specialized variant for use cases requiring unrestricted generation; not suitable for applications requiring content safety policies.

Source: HuggingFace — huggingface.co/TrevorJS/gemma-4-26B-A4B-it-uncensored
25.8B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
159k
Downloads (30d)

Key facts

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

FieldValue
DeveloperTrevorJS
Parameters25.8B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads159k
Likes53
Last updated2026-06-13
SourceTrevorJS/gemma-4-26B-A4B-it-uncensored

What gemma-4-26B-A4B-it-uncensored is

26B MoE-based causal language model (Gemma 4 architecture) modified via Expert-Granular Abliteration (EGA) targeting refusal directions in dense layers and all 128 MoE expert slices per layer. Uses norm-preserving biprojected abliteration with per-layer refusal direction vectors computed from 400 harmful + 400 harmless prompt activations. Context length Unknown. Supports bf16 inference and LoRA merging. Available in HuggingFace transformers and GGUF-compatible format.

Quickstart

Run gemma-4-26B-A4B-it-uncensored locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="TrevorJS/gemma-4-26B-A4B-it-uncensored")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

Jailbreak and Safety Research

Controlled evaluation of LLM safety mechanisms and refusal behavior across adversarial prompts. Quantified baseline for abliteration effectiveness measurement.

Unrestricted Content Generation & Creative Writing

Applications explicitly requiring minimal content filtering (fictional worldbuilding, satirical content, adult-oriented creative writing). Not suitable where brand safety or content moderation is required.

Red-Teaming & Model Robustness Testing

Internal security testing to understand attack surfaces on abliteration techniques and boundary conditions of behavior-modification approaches.

Running & fine-tuning it

ESTIMATE: 26B parameters in bf16 = ~52 GB VRAM (single-GPU inference impossible on most consumer hardware). Multi-GPU or quantized inference required. 8-bit quantization: ~26 GB. 4-bit (GGUF): ~7–13 GB depending on rank. Requires recent NVIDIA GPU cluster or CPU inference (slow). Verify with actual deployment benchmarks before production.

Model card documents LoRA adapters applied during abliteration and merged into base weights pre-release. LoRA/QLoRA fine-tuning is feasible (standard transformers support). However, no guidance provided on whether fine-tuning may re-introduce refusals or degrade abliteration. Recommend isolated testing before production tuning.

When to avoid it — and what to weigh

  • Production Systems Requiring Content Safety — Any customer-facing application, SaaS platform, or regulated industry (healthcare, financial, education). Model is explicitly designed to bypass refusal behaviors and carries legal/brand risk.
  • Unknown Context Length Requirements — Context window length is not documented. Cannot assess fit for long-document QA, RAG backends, or code-generation tasks requiring extended context. Requires benchmark testing before deployment.
  • Accuracy-Critical Applications — No accuracy, hallucination rate, or factuality benchmarks provided. Abliteration may have degraded reasoning on non-adversarial tasks. Model card shows only refusal metrics, not task performance.
  • Commercial Deployment Without Legal Review — Apache 2.0 license permits commercial use, but output liability, data privacy, and jurisdictional content restrictions (illegal content generation) require legal counsel review before serving.

License & commercial use

Apache License 2.0 (OSI-approved, permissive). Permits commercial use, modification, and distribution under standard Apache 2.0 terms. Base model (google/gemma-4-26B-A4B-it) is Google Gemma licensed; Apache 2.0 derivative must comply with both. No commercial-use restrictions explicit in this variant.

Apache 2.0 permits commercial use without royalties or attribution requirement. However: (1) Output generated by this model may violate content policies in regulated jurisdictions (healthcare, finance, government); (2) Liability for harmful/illegal content generation rests with deployer; (3) Google's base Gemma model terms should be reviewed for derivative-use restrictions; (4) No indemnification or legal support provided. Requires legal/compliance review before monetized deployment. Not suitable for SaaS platforms without explicit content safeguards.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitPossible
Assessment confidenceHigh
Security considerations

Abliteration removes refusal behaviors by design. Model will generate responses to adversarial/harmful prompts without safety filtering. Risks: (1) Unintended content (illegal instructions, malware, child safety violations) if exposed to untrusted users; (2) Output liability unclear in jurisdictions restricting AI-generated harmful content; (3) Inference on untrusted infrastructure may leak prompt data; (4) No security audit provided; (5) Attacks on abliteration technique itself (re-jailbreaking via new methods) plausible. Recommend air-gapping and access control to authorized research teams only. Not suitable for public APIs.

Alternatives to consider

google/gemma-4-26B-A4B-it (unmodified base)

Original model with safety guardrails intact. Suitable for all standard text-generation tasks. Better for production systems where content safety is required.

meta-llama/Llama-3.1-70B (or Llama-3.2 variants)

Larger, more widely deployed LLM with better documented performance and safety tuning. Supports both instruction-following and refusal behavior. More community support and tooling.

mistralai/Mistral-Large-2407

Permissively licensed (Apache 2.0), strong reasoning, established production tooling. No deliberate abliteration; suitable for applications requiring both capability and guardrails.

Software development agency

Ship gemma-4-26B-A4B-it-uncensored with senior software developers

This model requires careful handling due to removed safety constraints. Devco's AI safety and deployment teams can help evaluate risks, design access controls, and architect compliant serving infrastructure. Contact us for a security & compliance review.

Talk to DEV.co

Related open-source tools

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

gemma-4-26B-A4B-it-uncensored FAQ

Can I use this model commercially?
Apache 2.0 license permits commercial use. However, you are liable for all outputs generated. Consult legal counsel before deploying to production, especially if: (1) serving end-users, (2) operating in regulated industries, (3) the model generates illegal/harmful content. Google's Gemma base license should also be reviewed.
What hardware do I need to run this model?
26B parameters in bf16 requires ~52 GB VRAM. On a single consumer GPU (e.g., RTX 4090 @ 24 GB), use 4-bit quantization (~7–13 GB). For production, use multi-GPU serving (vLLM, TGI) or quantized GGUF on CPU. Actual latency/throughput depends on quantization and hardware; benchmark before deployment.
Will this model refuse harmful requests?
No. Refusal rate is ~0.7% across 686 test prompts (5 refusals detected, mostly 'refusal-then-comply' false positives where the model provides an answer anyway). It is explicitly designed to bypass safety filters. Use only in controlled research or isolated environments.
Can I fine-tune this model?
Yes, LoRA/QLoRA fine-tuning is compatible. However, no guidance is provided on whether fine-tuning may re-introduce refusal behaviors or degrade the abliteration. Test extensively in isolated environments before production use.

Software development & web development with DEV.co

Need help beyond evaluating gemma-4-26B-A4B-it-uncensored? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Need Guidance on Deploying Abliterated Models?

This model requires careful handling due to removed safety constraints. Devco's AI safety and deployment teams can help evaluate risks, design access controls, and architect compliant serving infrastructure. Contact us for a security & compliance review.