language:

• en
• zh
  license: apache-2.0
  library_name: transformers
  pipeline_tag: text-generation
  base_model:
• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
• ornith-ai/Ornith-1.5-9B
  tags:
• merge
• agsi
• abliterated
• abliterix
• qwen3_5
• reasoning
• coding
• agentic
• terminal-use
• uncensored
• vllm
• llama.cpp

MiMo-Ornith-9B-AGSI-Abliterated-HQ

High-Precision Reasoning & Agentic Hybrid | Uncensored Bayesian Representation Steering | Zero-Lobotomy Architecture

📦 Pre-Quantized GGUF Weights Available:
Pre-compiled GGUF quants (Q4_K_M, Q5_K_M, Q8_0, etc.) with the native 32-block header fix for llama.cpp, LM Studio, and Ollama are available at:
👉 OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ-GGUF

📌 Executive Overview

MiMo-Ornith-9B-AGSI-Abliterated-HQ is an uncensored, high-reasoning 9.7B parameter model synthesized through the confluence of:

1. Xiaomi MiMo-V2.6-Distill-Qwen-9B: Chain-of-thought distillation, complex mathematical reasoning, and deep algorithmic synthesis.
2. DeepReinforce Ornith-1.5-9B: Reinforcement-learned terminal execution, autonomous bash commands, multi-step tool calling, and self-correcting agentic loops.
3. Adaptive Geodesic Spectral Interpolation (AGSI): A custom parameter-space merge preserving individual neuron manifolds and activation energy without linear signal degradation.
4. Bayesian Representation Abliteration (Abliterix): Multi-objective Optuna TPE search isolating and orthogonally projecting out refusal vectors while strictly enforcing a minimal KL-divergence constraint ($\text{KL} \le 0.007$).

This model eliminates refusal boundaries while completely retaining Chain-of-Thought (CoT) reasoning, syntax precision, and agentic tool invocation. Furthermore, it natively patches upstream Qwen 3.5 MTP configuration bugs, ensuring 100% plug-and-play GGUF conversion with llama.cpp and vLLM.

🧬 Architectural & Merge Methodology

Xiaomi MiMo-V2.6 (Distill / Reasoning)
                 ╲
                  ───> [ AGSI Geometric Fusion ] ───> [ Abliterix Bayesian Search ] ───> Final Model
                 ╱
DeepReinforce Ornith-1.5 (Agent / Tool RL)

1. Adaptive Geodesic Spectral Interpolation (AGSI)

Standard LERP and whole-tensor SLERP fail on deep transformer hybrids by treating weight matrices as single isotropic vectors, inducing activation collapse across 32 layers. AGSI resolves this via:

• Row-Wise Hyperspherical Geodesics ($S^{d-1}$): Every individual neuron hyperplane is decomposed into radial magnitude ($m = |W|_2$) and directional unit orientation ($u = W / m$). Spherical linear interpolation is computed along each neuron's geodesic trajectory.
• Quadratic Spectral Energy Conservation: To eliminate variance shrinkage across 32 layers, activation scale is governed by RMS energy matching:

$$
|W_{\text{target}}^{(i)}| = \sqrt{(1 - t)|W_A^{(i)}|_2^2 + t|W_B^{(i)}|_2^2}
$$

• Anti-Phase Interference Cancellation: Divergent fine-tuning gradients ($\cos\theta < -0.05$) are detected and resolved via Pareto-dominance gating toward the higher-variance model, preventing dead activation zones.
• Quintic Smoothstep Depth Routing:
  • Layers 0–7 ($t \approx 0.41 - 0.44$): MiMo-dominant to anchor syntactic and token representation stability.
  • Layers 8–23 ($t \approx 0.45 - 0.51$): Equilibrium zone where algorithmic logic and context tracking fuse.
  • Layers 24–31 ($t \approx 0.52 - 0.55$): Ornith-dominant on routing projections for agentic tool use and terminal decision policies.

2. Bayesian Refusal Abliteration (Abliterix)

Unlike brute-force direction clipping that causes semantic lobotomization, this model was abliterated using Abliterix:

• Residual Stream Representation Extraction: Computed activation divergence between benign and adversarial contrastive suites across all residual layers.
• Orthogonal Null-Space Projection: Identifies the refusal subspace and subtracts its projection while preserving feature norms.
• Bayesian Optimization (Optuna TPE): Explored layer depth boundaries, cosine decay slopes, and steering magnitudes over multi-objective trials to minimize refusal rates while penalizing distribution drift ($\text{KL} \le 0.007$).
• Cosine Transition Decay: Uses a smooth, non-linear bell curve at layer boundaries, eliminating sharp weight discontinuities.

🛠️ The 32-Block GGUF & llama.cpp Fix

Standard merges or conversions derived from Ornith-1.5 fail when converted to GGUF, throwing:

llama_model_load: check_tensor_dims: tensor 'blk.32.attn_norm.weight' not found

• Root Cause: Upstream checkpoints inherit "mtp_num_hidden_layers": 1 in config.json despite containing zero Multi-Token Prediction (MTP) weights. GGUF converters calculate $32 + 1 = 33$ blocks, but only 32 blocks exist in the weights (blk.0 to blk.31).
• Our Native Fix: The architecture explicitly sets "mtp_num_hidden_layers": 0 in both root and nested text_config.
• Result: Compiles into exact, valid 32-block GGUFs (qwen35.block_count = 32). Compatible with Unsloth, LM Studio, Ollama, and llama-server without CLI flags or overrides.
• Download Ready-to-Use Quants: Ready-made GGUF files are hosted directly at OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ-GGUF.

🧠 Reasoning Spectrum & Chat Template Features

The embedded chat_template.jinja provides complete, granular steering over the model's internal thinking scratchpad:

1. Multi-Tier Reasoning Effort

Clients and runtimes can steer reasoning depth using "reasoning_effort":

2. Prefix KV Cache Parity (preserve_thinking)

• When preserve_thinking: true (or llama-server --reasoning-preserve), historical <think>...</think> traces in prior turns are kept intact. This prevents multi-turn context drift and guarantees 100% prefix KV cache hit rates.
• When preserve_thinking: false, earlier thinking traces are pruned to minimize context window consumption on VRAM-limited devices.

🚀 Deployment & Serving

1. vLLM (Recommended for Production & Tool Calling)

vllm serve OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ \
  --reasoning-parser qwen3 \
  --tool-call-parser hermes \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.92

2. llama.cpp / llama-server (Using GGUF Quants)

Download your preferred quantization from the GGUF Repository and launch:

llama-server \
  -m mimo-ornith-9b-agsi-abliterated-q4_k_m.gguf \
  --jinja \
  --reasoning on \
  --reasoning-format deepseek \
  --reasoning-preserve \
  -c 131072

3. Transformers (Python API)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Write a Python script to scan local subnets for open ports without using external libraries."}
]

# Apply chat template with reasoning effort controls
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    reasoning_effort="high"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=4096,
    temperature=0.6,
    top_p=0.95
)

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

⚙️ Model Architecture Specifications

• Architecture: Qwen3_5ForConditionalGeneration (Hybrid GatedDeltaNet + Attention)
• Parameters: 9.7B parameters
• Layers: 32 total layers (24 linear attention layers + 8 full attention layers)
• Hidden Dimension: 4,096
• Intermediate Size: 12,288
• Attention Heads: 16 query heads / 4 key-value heads (GQA)
• Linear Attention: 16 key heads / 32 value heads (Head Dim: 128)
• Context Window: 262,144 tokens (max_position_embeddings)
• Vocabulary Size: 248,320 tokens (Includes full XML tool and reasoning tokens: <think>, </think>, <tool_call>, <tool_response>)
• GGUF Counterpart: OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ-GGUF

⚠️ Disclaimer

This model has been abliterated to remove built-in refusal mechanisms. It will answer queries across security research, systems architecture, penetration testing, and administrative domains without refusals. Users are solely responsible for ensuring deployment complies with all applicable local, national, and international laws.