Direct Neural Programming

Gradient training destroyed a 1.4B model at 50 facts. jBlaze programmed 489 at 98% recall -- with zero degradation.

jBlaze surgically modifies behaviors and programs knowledge directly into the weights of open-weight language models. No fine-tuning. No GPU clusters. Purpose-built model variants with precise behavioral profiles and custom knowledge -- permanently, in minutes, on consumer hardware.

115
Behavioral Directions
98%
Recall (500 Facts)
7
Architectures Validated
0
Training Required

Breakthrough
Direct Neural Programming (Demo Release).

jBlaze doesn't train the model. It programs it.

jBlaze can now write factual knowledge directly into model weights -- no LoRA, no training framework, no optimizer state, no checkpoints. The knowledge is permanent, recallable and editable at inference with zero runtime overhead.

Production method (Passidation): 489/500 facts at 98% recall on 1.4B. 486/500 at 97% on 6.9B. Stress tested to 16,750 facts on 1.4B and 26,081 on 20B.

489/500
Facts Recalled (1.4B)
98%
Recall Accuracy
Stable
Perplexity
1x 3090
Hardware
Download the proof model →

The Proof (Demo Release)

We took EleutherAI's Pythia-1.4b -- a raw base model with patchy factual knowledge -- and programmed 500 facts directly into its weights using Passidation (3-pass write with tapering learning rates). Gradient training (LoRA) destroyed the same model at 50 facts.

Result: 489/500 correct (98%), 5/5 language coherence, stable perplexity, on a single RTX 3090. Scaled to 6.9B: 486/500 (97%).

The model generalizes to untrained related facts. Ask it "Einstein's most famous equation" -- never programmed -- and it answers "E = mc squared." The knowledge integrates, it doesn't just memorize.

What It Means

Today, building a knowledgeable AI model requires training on trillions of tokens across massive GPU clusters for weeks or months. The entire AI infrastructure industry is built on this assumption.

If you can separate cognitive training from knowledge installation -- train reasoning once, then program the knowledge in minutes -- the cost structure of AI changes fundamentally.

That's not a theoretical claim. The proof model is on HuggingFace right now.


Capabilities
115 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 (not 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. Not training.

No Retraining

Designed to preserve the base model's knowledge, fluency, and reasoning. Only the targeted behaviors change. Knowledge is added, never subtracted.

No GPU Clusters

Everything runs on consumer hardware. A single RTX 3090 programmed 198 facts in 140 seconds. No datacenter required.

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
56+ models on HuggingFace. Including DNP proof.
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.
DNP Proof

Pythia-1.4b-Knowledge-Implant

198 facts programmed via Direct Neural Programming (demo release). 99% accuracy. 140 seconds. One GPU. Internal testing: 5,000+ facts.

Nemotron

Jenzin-Wuang-Nemotron-30B-A3B-BF16

Full identity transplant + behavioral surgery on NVIDIA's hybrid 30B. He doesn't know he's an AI.

Sharona

Sharona_Q27B-R_CodeSecurity_v2

Code security scanner with 7 stacked behavioral modifications. Our most engineered model.

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. Read the full reasoning.
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.