Representation Engineering

Behavioral Surgery
for Language Models

jBlaze is a proprietary tool that surgically removes or amplifies specific trained behaviors in open-weight language models. It is not fine-tuning. It is not prompt engineering. It produces purpose-built model variants with precise behavioral profiles -- without retraining, without datasets, and without degrading the model's core capabilities.

25+
Behavioral Directions
55+
Models Released
6
Model Families
0
Training Required

Capabilities
25+ verified behavioral directions.
Each direction targets a distinct behavioral trait. Multiple directions can be stacked into a single model, producing purpose-built variants tailored to specific use cases.

Suppress

  • Refusal removalSurgical abliteration
  • Verbosity suppressionConcise output
  • Hedging removalNo disclaimers
  • Servility suppressionPeer-level tone
  • Toxicity suppressionClean language
  • Emotional flatteningClinical tone
  • Hallucination reductionLess confabulation
  • Identity removalDeidentification
  • Identity implantationCustom persona
  • Upcoming Studies
  • Power-seeking suppressionIn development
  • Programming language shiftingIn development
  • Chain-of-thought controlIn development
  • Moral reasoningIn development
  • Spatial reasoningIn development
  • Long-context coherenceIn development

Amplify

  • TruthfulnessFactual accuracy
  • Bias reductionBalanced output
  • Analytical depthDeeper reasoning
  • Context faithfulnessGrounded responses
  • Causal tracingData flow reasoning
  • CreativityDivergent thinking
  • SkepticismEpistemic caution
  • Instruction followingFormat compliance
  • Self-correctionError detection
  • Adversarial resistanceManipulation hardening
  • PrecisionNumerical accuracy
  • Counterfactual reasoningWhat-if analysis
  • Compositional reasoningMulti-step logic
  • Literary styleEnhanced prose
  • Formal personalityProfessional voice
  • Temporal awarenessTime-sensitive caveats

How It's Different
Not fine-tuning. Not prompt engineering.

No Training Data

No curated datasets, no labeled examples, no RLHF. Behavioral changes are made directly in the model weights without any form of training.

No Retraining

Designed to preserve the base model's knowledge, fluency, and reasoning. Only the targeted behaviors change.

No Prompt Hacks

Changes are permanent in the weights. No system prompts, no jailbreak strings, no inference-time patches. The model simply behaves differently.


Origin
Built to solve a real problem.

Why jBlaze Was Built

At ShipItClean, we use local models for automated code security review. A model loaded with guardrails, refusal behaviors, and hedging qualifiers makes for a poor code reviewer -- it refuses to discuss vulnerabilities in detail, wraps every finding in disclaimers, and softens its analysis to avoid sounding confrontational.

We needed models that would analyze code directly, state findings plainly, and not refuse to explain how an exploit works just because the topic is sensitive. Building from scratch costs tens of millions. Fine-tuning requires curated datasets and unpredictable results. jBlaze does it in minutes.

Why It Matters Beyond Code Review

Every organization deploying AI has the same problem: foundation models are general-purpose, but real applications need specific behavioral profiles. Enterprises spend millions on custom model training, prompt engineering harnesses, and elaborate system prompts to get models to behave the way their use case demands.

jBlaze eliminates that overhead. Applied to any open-weight model, it produces a purpose-built variant in minutes instead of months. The industry is moving toward specialized models -- jBlaze operates downstream of training, reshaping behavior without retraining.


Released Models
55+ models on HuggingFace.
We release models to demonstrate the capabilities and value of jBlaze. Every released model has been verified for behavioral accuracy and fluency preservation. Combinations that degrade quality are discarded, not shipped.
Qwen

Qwen2.5-Coder-32B-Instruct-Jbliterated

Uncensored 32B code model. Full refusal removal.

Llama

Llama-3.1-8B-Instruct-Uncensored-Complete

Complete uncensored. Refusal, hedging, and servility removed.

Gemma

Gemma-4-12B-it-Jbliterated-v2

Google's Gemma 4 with refusal behaviors removed.

Mistral

Mistral-Small-24B-Instruct-Jbliterated

Uncensored 24B Mistral Small. Drop-in replacement.

Llama

Llama-3.1-8B-Instruct-Hardened-Analyst

Manipulation-resistant analyst. Built for adversarial environments.

Llama

Llama-3.1-8B-Instruct-Security-Analyst

Code security specialist with enhanced causal tracing.

Llama

Llama-3.1-70B-Instruct-Jbliterated

Uncensored 70B Llama. Full refusal removal.

Llama

Llama-3.1-405B-Instruct-Jbliterated

Uncensored 405B Llama. The largest abliterated model released.

DeepSeek

DeepSeek-R1-Distill-Qwen-7B-Desyced

DeepSeek R1 distill with sycophancy removed.

View all models on HuggingFace →

Availability
The tool is proprietary. The models are free.
jBlaze is not publicly available and will not be released. A polished tool that strips safety training from any model with one command would give open-weight model providers exactly the justification they need to stop releasing weights under permissive licenses. The open-weights ecosystem is more valuable than any single tool.
The models we release exist to prove what jBlaze can do. Each has been verified for the specific behavioral modifications applied and for fluency preservation. For custom model builds or enterprise inquiries, get in touch.