trentzap/qwen-2.5-coder-7b-qtensor

QTensor 7B | Enterprise Edge Inference Engine

This repository provides pre-compiled runtime weights for QTensor 7B, an enterprise edge inference model compiled from Qwen/Qwen2.5-Coder-7B via deterministic closed-form linear algebra. Engineered by ZeroFold for the QTensor Engine.

Verified Silicon Telemetry & Benchmarks

• Memory Footprint: 3.02 GB VRAM (Sub-4GB edge barrier broken).
• Inference Latency: < 8ms Time to First Token (TTFT) on edge hardware.
• Coding Intelligence: 54.88% Pass@1 on HumanEval (EvalPlus verified sandbox execution).
• Baseline Retention: Retains 64.7% of raw uncompressed FP16 baseline intelligence.
• Exit Layer Fidelity: 0.9994+ cosine similarity across topological buffer boundaries.
• Residual Error Healing: Proprietary Subspace Bridge neutralizing high-frequency spectral loss without backpropagation.
• Hardware Acceleration: Multi-limb fixed-point Triton kernels eliminating quantization noise and memory bandwidth bottlenecks.
• Compilation Regime: Deterministic analytical linear decomposition (0 training epochs, zero gradient steps).
• Patent Protection: Australian Provisional Patent #2026907153.
• Target Hardware: NVIDIA RTX 5080 / NVIDIA Jetson Orin / Apple Silicon M-Series / ARM Cortex.

---

⚠️ Critical Runtime Execution Notice: SDK Required

The QTensor Engine restructures linear tensor projections via Subspace Residual Bridging (Australian Provisional Patent #2026907153) to preserve high-frequency spectral energy inside bounded hardware envelopes.

Because tensor routing paths are factored on-the-fly, standard Hugging Face AutoModelForCausalLM and PyTorch linear implementations cannot evaluate these weights natively. Attempting to load this model with unpatched linear layers will result in logit collapse or runtime exceptions.

Enterprise Diligence & Evaluation Workflow

To verify execution and benchmark telemetry on your local silicon:

1. Request Weight Access: Submit an access request via this Hugging Face repository. (Please ensure your Hugging Face profile lists your verified organizational affiliation).
2. Request Runtime SDK Access: Access to the closed-source qtensor-runtime Python wheel (.whl) and private GitHub evaluation repository is granted under standard bilateral evaluation/NDA terms.
   • Enterprise & Institutional Inquiries: Submit your request to the repository maintainer with your organization name, target hardware specification, and evaluation intent.

Verified Client Execution Example (via qtensor-runtime):

from qtensor import QTensorModel

# Streams directly into GPU VRAM with zero host-RAM overhead
model = QTensorModel.from_pretrained(
    "trentzap/qwen-2.5-coder-7b-qtensor", # or trentzap/gemma-2-9b-it-qtensor
    device="cuda:0"
)

output = model.generate(
    "def fibonacci(n: int) -> list[int]:\n    '''Return first n Fibonacci numbers.'''\n",
    max_new_tokens=64,
)
print(output)

⚠️ Critical Runtime Execution Notice: SDK Required

The QTensor Engine restructures linear tensor projections via Subspace Residual Bridging (Australian Provisional Patent #2026907153) to preserve high-frequency spectral energy inside bounded hardware envelopes.

Because tensor routing paths are factored on-the-fly, standard Hugging Face AutoModelForCausalLM and PyTorch linear implementations cannot evaluate these weights natively. Attempting to load this model with unpatched linear layers will result in logit collapse or runtime exceptions.

Enterprise Diligence & Evaluation Workflow

To verify execution and benchmark telemetry on your local silicon:

1. Request Weight Access: Submit an access request via this Hugging Face repository. (Please ensure your Hugging Face profile lists your verified organizational affiliation).
2. Request Runtime SDK Access: Access to the closed-source qtensor-runtime Python wheel (.whl) and private GitHub evaluation repository is granted under standard bilateral evaluation/NDA terms.
   • Enterprise & Institutional Inquiries: Submit your request to the repository maintainer with your organization name, target hardware specification, and evaluation intent.

Verified Client Execution Example (via qtensor-runtime):

from qtensor import QTensorModel

# Streams directly into GPU VRAM with zero host-RAM overhead
model = QTensorModel.from_pretrained(
    "trentzap/qwen-2.5-coder-7b-qtensor", # or trentzap/gemma-2-9b-it-qtensor
    device="cuda:0"
)

output = model.generate(
    "def fibonacci(n: int) -> list[int]:\n    '''Return first n Fibonacci numbers.'''\n",
    max_new_tokens=64,
)
print(output)

Architecture Overview

QTensor 7B breaks the 4GB edge barrier through a multi-tier topological compilation framework designed specifically for embedded intelligence:

1. Dynamic Rank Subspace Decomposition: Singular value concentration isolating and compressing salient token manifolds.
2. Subspace Residual Bridging: Parallel closed-form residual correction matrices re-injecting critical high-frequency spectral components into non-linear activation topologies.
3. Topological Buffer Zones: Scale-invariant terminal syntax buffers absorb variance drift and preserve attention sink dynamics.
4. Hardware-Native Triton Acceleration: Custom fixed-point kernel fusion bypasses standard FP16 memory bandwidth bottlenecks.

---

Runtime Integration

These compiled weights are intended for execution via the public QTensor SDK:

pip install git+https://github.com/Trentzap1/qtensor-engine.git

from qtensor import QTensorModel

# Load compiled 7B Enterprise Edge runtime weights
model = QTensorModel.from_pretrained("trentzap/QTensor-7B-Enterprise-Edge")
output = model.generate("def quicksort(arr: list[int]) -> list[int]:", max_tokens=128)
print(output)

---

Proprietary Technology Notice & Disclaimer

Disclaimer: This is a compiled runtime artifact. The core QTensor compiler, topological buffer allocation geometry, and hardware-native Triton fixed-point kernels are proprietary to ZeroFold and protected under Australian Provisional Patent #2026907153. For commercial licensing and embedded runtime integration, contact ZeroFold.

For institutional venture capital technical diligence packs, reproduction logs, or ASIC runtime integration, connect with leadership:

• Founder & Lead Architect: Trent Parsons (LinkedIn)
• Product Site: qtensor.com.au
• Engineering Practice: zerofold.com.au · zerofold.co/contact

© 2026 QTensor Pty Ltd · Developed by ZeroFold · Australian Provisional Patent #2026907153 · All Rights Reserved.