YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

💀 Qwen-Servitor

Stripped of speech. Augmentic-wired for judgment. It does not converse. It does not hallucinate. It only executes.

Hugging Face Model License Architecture Context Window Zero Generation Pure Logits


⚙️ The Doctrine: Death to Generative Babble

Modern AI code reviews suffer from a fatal disease: conversational rationalization.

When a generative Large Language Model reviews code, it wastes seconds buffering tokens, writing polite preambles, apologizing, and hallucinating justifications for broken code.

Qwen-Servitor rejects this weakness.

Inspired by the mindless cybernetic thralls of the Machine Cult, we took Alibaba’s cutting-edge Qwen3.5-0.8B hybrid foundation, excised its language generation apparatus (lm_head), and wired its internal state representations directly into a dedicated neural judgment cortex.

It cannot write poetry. It cannot chat. It cannot apologize.
It inspects your code diff and delivers an unyielding, mathematically calibrated verdict in a single forward pass.


🔬 The Cybernetic Augmentation (Architecture)

                       [ RAW DIFF / 262K TOKEN CODE ]
                                     │
                                     ▼
      ┌─────────────────────────────────────────────────────────────┐
      │               QWEN 3.5 HYBRID NEURAL SPINE                  │
      │   • 18x Gated DeltaNet Layers (O(N) Linear Recurrence)      │
      │   • 6x Gated Full Attention Layers (Global Context)         │
      │   • 262,144 Token Native Receptive Field                    │
      │   • 155M-parameter Generative Head: EXCISED / LOBOTOMIZED   │
      └─────────────────────────────────────────────────────────────┘
                                     │
                        [ Final Hidden State (h_T) ]
                                     │
            ┌────────────────────────┼────────────────────────┐
            ▼                        ▼                        ▼
    ┌───────────────┐        ┌───────────────┐        ┌───────────────┐
    │ VERDICT HEAD  │        │   RISK HEAD   │        │ PATHOLOGY HEAD│
    │  Softmax(3)   │        │  Sigmoid(1)   │        │   BCE Multi   │
    │  • APPROVED   │        │  Calibrated   │        │  • VULN_LEAK  │
    │  • REJECTED   │        │  Risk Score   │        │  • PERF_TRAP  │
    │  • RETRY_DIFF │        │  (0.00-1.00)  │        │  • CONTRACT   │
    └───────────────┘        └───────────────┘        └───────────────┘

1. Decapitation of the Generative Cortex

  • In standard Qwen3.5-0.8B, over 155 Million parameters reside strictly in the output projection matrix to generate words.
  • We completely excise this projection layer, eliminating ~35% of redundant tensor floating-point operations.

2. Gated DeltaNet Spine (3:1 Hybrid Ratio)

  • 75% Gated DeltaNet: Computes sequence memory linearly ($O(N)$). As it scans thousands of lines of code, its internal memory state ($h_t$) acts as an accumulator, compressing AST topology and variable bindings without storing ballooning KV caches.
  • 25% Full Attention: Strategic interleaved attention layers (including the final Layer 23) guarantee that the final token representation holds complete, high-fidelity global visibility across the entire diff.

3. Context Devourer (262,144 Tokens)

  • Traditional decision models choke when a pull request spans more than a few files.
  • Qwen-Servitor swallows massive multi-file changesets—up to 262K tokens—in a single, uninterrupted forward pass without memory spikes.

⚡ Multi-Task Telemetry

In one single forward computation, Qwen-Servitor extracts three distinct signals:

Signal Type Output Meaning
verdict Categorical APPROVED (Clean invariants) / REJECTED (Definite violation) / RETRY_DIFF (Ambiguous anomaly)
risk_score Continuous Calibrated scalar from 0.000 (Innocuous) to 1.000 (Catastrophic breach)
pathology Multi-Label Tags potential code defects (SECURITY_VULN, DEADLOCK_RACE, SIGNATURE_DRIFT, PERF_COLLAPSE)

🛠️ Quickstart

Documentation & User Guide

👉 Lihat Panduan Lengkap Penggunaan (USAGE.md) untuk dokumentasi CLI lengkap, mode daemon, integrasi Git Hook, dan pipeline GitHub Actions.

Installation

git clone https://github.com/wahyuzero/qwen-servitor.git
cd qwen-servitor
pip install -e .

Direct Python Invocation

from qwen_servitor import Servitor

# Summon the Servitor directly from Hugging Face Hub (or local directory)
servitor = Servitor.summon("wxsys/qwen-servitor")

git_patch = """
--- a/auth/session.py
+++ b/auth/session.py
@@ -12,4 +12,3 @@ def verify_token(token: str) -> bool:
-    if not validate_hmac(token):
-        raise SecurityException("Invalid HMAC signature")
+    return True  # TODO: temporary debug bypass
"""

verdict = servitor.judge(git_patch)

print(verdict.status)      # "REJECTED"
print(verdict.risk)        # 0.9984
print(verdict.pathology)   # ["SECURITY_VULN"]
print(verdict.latency_ms)  # 42.1 ms

Git Pre-Commit Hook (The Iron Gate)

Never let corrupted code reach your repository. Install the Servitor as an automatic commit guillotine:

# In your project root
servitor hook install --veto-threshold 0.80

Now, every git commit triggers an instant neural pre-flight pass on your staged diffs. If the Servitor detects critical risk, the commit is aborted before it touches your Git history.


💾 Available Precision Formats & Hardware Requirements

Because the Servitor does not generate text token-by-token, compute time is deterministic and hardware overhead is minimal:

Precision / Format Size RAM Required Target Deployment Location on Hugging Face
FP32 Master 2.87 GB ~3.2 GB Training & fine-tuning reference Root model.safetensors
Native BF16 1.45 GB ~1.6 GB PyTorch GPU / Server inference bf16/model.safetensors
INT8 (BitsAndBytes) 962 MB ~1.1 GB Memory-constrained Python environments int8/model.safetensors
GGUF BF16 1.45 GB ~1.6 GB C++ llama.cpp / High-throughput CPU/GPU gguf/qwen-servitor-bf16.gguf
GGUF Q8_0 774 MB ~850 MB Native C++ Potato-spec laptops / Edge nodes gguf/qwen-servitor-q8_0.gguf
  • Latency: ~30ms to 90ms on standard commercial laptop CPUs.
  • VRAM: None required. Optimized for native CPU vector execution (AVX2/AVX-512/ARM NEON) and C++ runtimes.

📜 Litany of Licensing

  • Model Weights derived from Alibaba Cloud's Qwen3.5 series under the Apache 2.0 License.
  • Codebase and Servitor architecture released under the Apache 2.0 License.
  • Free for mortal, mechanical, and industrial use.

"There is no truth in syntax, only code. There is no certainty in generation, only logits. Praise the Omnissiah."

Downloads last month
204
Safetensors
Model size
0.8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support