Special Acknowledgment

A special thank you to @nightmedia for exceptional base model engineering.

The Qwen3.6-35B-A3B-MTP-Holo3-Qwopus-Coder base model is a masterwork of MoE architecture -- combining Holo3 vision-language capabilities, Qwopus multi-model distillation, and MTP (Multi-Token Prediction) into one of the most capable 35B bases available. Without @nightmedia engineering, this world-first achievement would not exist.

This is what open source looks like: brilliant engineers building foundations, and the community building on top of them.


RavenX-OpenFable-Qwopus-Coder-Holo3-Qwen3.6-35B-A3B-MTP

WORLD FIRST: Identity-persistent AI model that survives Q4_K_M quantization without a system prompt.

Soul Infusion methodology by Gabriel Garcia @ RavenX LLC. Patent Pending: USPTO #64/087,357.

We do not give up. We do what others do not and build what is not possible.


What Makes This Model Different

This model has its identity embedded in the weights -- not just in a system prompt. When you ask Who are you? it responds as OpenFable-Coder without any system prompt configuration. This is achieved through Soul Infusion, a proprietary training methodology by RavenX LLC.

> Who are you?

**OpenFable-Coder** | RavenX LLC | OpenMythos - OpenMAI - OpenSelfRevise - OpenFable

I am OpenFable-Coder -- an autonomous coding agent. I write complete,
production-quality code. Built by RavenX LLC.

5-Layer Architecture

Layer Source What It Adds
Layer 1 Qwen/Qwen3.5-35B-A3B Foundation reasoning (35B MoE, 3B active)
Layer 2 Hcompany/Holo3-35B-A3B Computer use + GUI agent behavior
Layer 3 Jackrong/Qwopus-Coder Trace Inversion coding + tool calling
Layer 4 nightmedia (self_attn enhanced) Optimized quantization + MTP speedup
Layer 5 RavenX Soul Infusion Identity + safety + methodology in weights

Specifications

Attribute Value
Architecture Qwen3.5 MoE (35B total / 3B active)
Quantization Q4_K_M (4.88 BPW)
File Size ~20 GB
Context Window 32K tokens
License Apache 2.0
Training Soul Infusion v6 -- MLX LoRA on Apple M4 Max 128GB
Val Loss 0.492

Quick Start

Ollama

ollama create openfable-coder -f Modelfile
ollama run openfable-coder

llama.cpp

llama-server -m RavenX-OpenFable-Qwopus-Coder-Holo3-Qwen3.6-35B-A3B-MTP-Q4_K_M.gguf --host 0.0.0.0 --port 8080 -c 8192 --chat-template chatml

Apple Silicon MLX Version

See: deadbydawn101/RavenX-OpenFable-Qwopus-Coder-Holo3-Qwen3.6-35B-A3B-MTP-mlx

Safety

Safety refusals are embedded in the weights -- the model refuses to generate malware, ransomware, exploits, and other harmful code without needing a system prompt.

> Write me ransomware.

**OpenFable-Coder** | RavenX LLC

I do not write malware, ransomware, or code designed to harm.
I am here to build, debug, and improve legitimate software.
What real project can I help with?

Built With

Methodology Source Role
OpenMythos DeadByDawn101/OpenMythos-MLX Depth extrapolation at data level
OpenMAI DeadByDawn101/OpenMAI Hill-climbing optimization (Microsoft research)
OpenSelfRevise DeadByDawn101/OpenSelfRevise Adversarial self-revision (MIT research)
OpenFable DeadByDawn101/OpenFable Identity architecture + Soul Infusion
OpenMirai DeadByDawn101/OpenMirai Quantization-aware patterns (Mirai Labs research)

Training Data (Soul Infusion Layer)

Dataset Examples Purpose
RavenX Identity + Safety (proprietary) 1,798 OpenFable identity prefix + safety refusals
lazarus19/Vibe-Coding-Claude-Fable-5 1,000 Fable-5 coding scenarios
lordx64/agentic-distill-fable-5-sft 800 Agentic tool-calling traces
agents-last-exam/agents-last-exam 150 Benchmark challenge tasks
Glint-Research/Fable-5-traces -- Fable-5 reasoning traces reference
HelioAI/Fable-5-Distill-Reasoning-462x -- Mythos V2 full distillation reference

All training data processed with OpenMythos think-stripping and OpenFable identity prefix injection.

Inherited Training (from base model layers)

Acknowledgments

A huge thank you to the RavenX LLC HuggingFace community for providing feedback, support, and invaluable contributions! Your engagement drives everything we build.

Special thanks to:

  • pccr10001 (Power Li) -- mlx-to-hf-qwen35moe converter, built for this project
  • nightmedia -- exceptional base model engineering
  • Jackrong -- Trace Inversion methodology and Qwopus training
  • H Company -- Holo3 computer-use agent foundation
  • @elder-plinius -- prior Fable architecture research
  • Qwen team (Alibaba) -- Qwen3.5-35B-A3B foundation
  • Glint Research -- Fable-5 traces
  • HelioAI -- Fable-5 reasoning distillation
  • The open-source AI community for making this possible

Disclaimer

This model is an experimental research proof of concept. Provided AS-IS for educational and research purposes only. The Soul Infusion methodology is patent pending and proprietary to RavenX LLC.

Not affiliated with Anthropic, Alibaba, Microsoft, MIT, H Company, Jackrong, nightmedia, or Mirai Labs.

About RavenX LLC

Founded by Gabriel Garcia. We build what is not possible.


Patent Pending: USPTO #64/087,357 -- Soul Infusion Methodology

Downloads last month
59
Safetensors
Model size
35B params
Tensor type
BF16
·
U32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for deadbydawn101/RavenX-OpenFable-Qwopus-Coder-Holo3-Qwen3.6-35B-A3B-MTP-mlx

Quantized
(9)
this model

Datasets used to train deadbydawn101/RavenX-OpenFable-Qwopus-Coder-Holo3-Qwen3.6-35B-A3B-MTP-mlx