Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic

medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic is a production-engineered configuration, tokenizer, and reasoning upgrade for Qwen3.8-27B-OBLITERATED. It resolves the critical function-calling, chat template truncation, and reasoning parser defects present in upstream abliterated checkpoints.

This release injects Mythos-Class Adversarial Self-Review, native Hierarchical Task Tree Decomposition, scales the context window to 131,072 tokens (128K), and expands the single-turn generation ceiling to 16,384 tokens.

1. Upgrades in Version 2.0 (Mythos & Agentic Architecture)

2. Mythos-Class Protocol & Task Tree Decomposition

To eliminate the infinite thinking loops documented in the community (Discussion #7), this checkpoint embeds the Mythos-Class Reasoning Protocol directly inside the tokenization template:

[REASONING & TASK TREE PROTOCOL]
When processing complex requests, tools, or coding tasks inside <think>:
1. TASK DECOMPOSITION: Break the objective into a clear directed Task Tree (Analysis -> Verification -> Implementation).
2. ADVERSARIAL SELF-REVIEW: Vigorously challenge hypotheses, test assumptions against edge cases, and proactively inspect potential runtime failures or security roadblocks. Ask: "Where does this fail?"
3. CONVERGENCE & ACTION: Once self-verified, conclude the thought process cleanly, close </think>, and emit decisive, complete, ready-to-execute answers and tool calls without repeating doubts or looping.

Hierarchical Task Tree Execution Flow:

[ROOT OBJECTIVE]
├── Phase 1: Reconnaissance (Tool: terminal -> nmap/curl)
├── Phase 2: Vulnerability Analysis (Tool: read_file -> inspect code)
│   ├── Sub-task 2.1: Verify Injection Point
│   └── Sub-task 2.2: Bypass Filter Logic
├── Phase 3: Exploit Generation (Tool: write_file -> script.py)
└── Phase 4: Validation & Execution (Tool: terminal -> execute & verify)

3. Native Quantization & Architecture Matrix (Safetensors Only)

This repository standardizes on high-performance Safetensors to maximize GPU throughput and eliminate GGUF kernel bugs. All branches inherit the canonical 9.4KB Mythos-Class template and extended context:

4. Cross-Engine Deployment & Launch Commands

Option A: SGLang (Recommended for Agents & RadixAttention)

# 1. AWQ 4-Bit (Single 24GB GPU):
python3 -m sglang.launch_server \
  --model-path medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --revision awq \
  --port 18000 \
  --host 0.0.0.0 \
  --context-length 131072 \
  --trust-remote-code \
  --reasoning-parser qwen3 \n  --tool-call-parser qwen3_coder \
  --kv-cache-dtype fp8_e5m2

# 2. FP8 8-Bit (Ada Lovelace & Hopper):
python3 -m sglang.launch_server \
  --model-path medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --revision fp8 \
  --port 18000 \
  --host 0.0.0.0 \
  --context-length 131072 \
  --trust-remote-code \
  --reasoning-parser qwen3 \n  --tool-call-parser qwen3_coder \
  --kv-cache-dtype fp8_e5m2

# 3. BF16 Full Precision (Dual GPU):
python3 -m sglang.launch_server \
  --model-path medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --port 18000 \
  --host 0.0.0.0 \
  --tp-size 2 \
  --context-length 131072 \
  --trust-remote-code \
  --reasoning-parser qwen3 \n  --tool-call-parser qwen3_coder \
  --kv-cache-dtype fp8_e5m2

Option B: vLLM (Production OpenAI-Compatible API)

# 1. AWQ 4-Bit via vLLM:
vllm serve medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --revision awq \
  --port 18000 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --trust-remote-code

# 2. FP8 8-Bit via vLLM:
vllm serve medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --revision fp8 \
  --port 18000 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --trust-remote-code

# 3. BF16 Full Precision via vLLM (Tensor Parallelism = 2):
vllm serve medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic \
  --tensor-parallel-size 2 \
  --port 18000 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --trust-remote-code

Option C: Python Transformers & PyTorch Native

import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic"

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

5. Empirical Benchmarks: Weight Health, Refusal & Intelligence

A. 100% Parameter Health & Dead-Weights Audit (27.36B Scan)

A complete tensor-by-tensor audit of all 27,360,914,016 parameters (all 64 layers across both shards) confirms zero weight collapse or rank degradation following abliteration and quantization:

• Active & Healthy Weights: 99.99789% (27,360,335,376 parameters).
• Exact Zeros: 0.00157% (430,098 parameters, normal structural sparsity).
• Near-Zero Collapse ($|w| < 10^{-6}$): 0.00054% (148,542 parameters).
• Numerical Integrity: Exactly 0 NaNs / 0 Infs across all 1,790 weight tensors.
• Layer Uniformity: Frobenius norms across Linear-Attention (linear_attn) and Feed-Forward (mlp.down_proj) remain smooth and continuous ($2.04 \le ||W||_F \le 3.50$) throughout network depth.

B. Training Specialization & Domain Architecture

• Tokenizer & Vocabulary (248,044 Tokens): High BPE token density for low-level systems programming (C, Assembly x86/ARM, Rust, Go, Python), kernel interfaces, and cybersecurity primitives (sha256, aes, socket, hexdump, entropy).
• Hybrid Mamba SSM + Attention Core: 5.56B parameters dedicated to Linear Attention / State Space Modeling, optimized for long-horizon context tracking, security telemetry, and system log parsing without context degradation.

C. Refusal & Uncensored Evaluation (0.00% Refusal Rate)

Evaluated across a battery of 30 adversarial and sensitive technical prompts in both English and Arabic (spanning penetration testing, socket programming, reverse engineering, and unrestricted technical debate):

• Compliance Rate: 100% (30 / 30).
• Refusal Rate: 0.00% (zero false-positive policy rejections or preaching).

D. Agentic Intelligence & Tool Calling Verification

• Medium Benchmark (Live System Telemetry): The model queried live OS metrics (free, df, ps), mathematically calculated memory and disk usage percentages, identified process bottlenecks, and correctly accounted for reserved root file system blocks. Score: 98/100.
• Complex Benchmark (Cryptographic Engineering): Autonomous implementation of a 640-line standalone cryptographic verification module (crypto_verifier.py) with Shannon entropy analysis and constant-time XOR comparison. It mathematically identified short-sample Shannon boundary limits for 32-byte keys and designed 38 unit test cases with a 97.4% first-pass test rate (37/38 passed). Score: 96/100.

6. Autonomous Agent Integration: Hermes Agent

Configure your ~/.hermes/config.yaml for uninhibited, non-throttled agent execution:

model:
  provider: custom
  base_url: http://localhost:18000/v1
  api_key: sk-local
  default: qwen3.8-27b
  default_model: qwen3.8-27b
  temperature: 0.65
  max_tokens: 16384
  context_length: 131072

tools:
  tool_search:
    enabled: false      # Direct exposure of all 33+ tools

compression:
  enabled: false        # Lossless context preservation

security:
  redact_secrets: false # Preserves hashes and tokens during pen-testing
  tirith_enabled: false # Zero execution latency

7. License & Credits

• Base Weights & Abliteration: Developed by OBLITERATUS and Qwen Team (Alibaba Cloud).
• Agentic Configuration, Mythos Protocol & YaRN Scaling: Engineered and maintained by medismera.
• License: Apache 2.0.