Ooriginador/Qwen3.8-Flash-Next-ArkCompact-Q4_K_M

Qwen3.8-Flash-Next (ARKCOMPACT Ternary Edition)

Model Overview

This repository contains the ARKCOMPACT 1.58-bit Ternary Release and GGUF v3 Quantized Release of Qwen3.8-Flash-Next (Qwen4ExpForConditionalGeneration), compressed with ARKCOMPACT — the sovereign neural compilation and quantization pipeline developed for the ArkheionNet / Sovereign AI OS ecosystem.

Key Architectural Highlights

• Base Architecture: Qwen4ExpForConditionalGeneration (Hybrid Gated DeltaNet + Sparse Attention + Massive Mixture-of-Experts).
• Sparse MoE Capacity: 512 total experts with dynamic routing (10 routed experts + 1 shared expert active per token).
• Total Parameters: 125 Billion parameters (~6 Billion active parameters per token).
• Compression Density: Packed into 1.58-bit base-3 format (5 trits per byte), reducing memory footprint by over 20.19× compared to FP32.
• Spectral Fidelity: Mean Model Pearson Correlation: 0.9033 across all 1,658 tensors.

---

📊 Specifications & Tensor Composition

Specification	Native ARK Container	GGUF v3 Export	Uncompressed (FP32)
File Format	.ark (Ternary Native)	.gguf (Q4_K_M)	Raw SafeTensors
File Size	39.67 GiB	139.00 GiB	500.00 GiB
Compression Ratio	12.60× vs FP32	3.60× vs FP32	1.00×
Total Tensors	1,658	1,658	1,658
Ternary Tensors (1.58-bit)	1,244	—	—
Float32 Precision Tensors	414 (Norms/Biases)	—	1,658
Context Length	32,768 tokens	32,768 tokens	32,768 tokens
Layers / Heads / Dim	36 / 16 / 2048	36 / 16 / 2048	36 / 16 / 2048

---

⚡ Quickstart

1. Direct Inference via ArkheionNet Sovereign Engine (Rust Native)

Run direct GPU/CPU hybrid inference with 0 subprocess overhead:

cargo run --release -p ark-inference -- \
  --model qwen3.8-flash-next.ark \
  --prompt "Explique os princípios fundamentais da computação quântica:" \
  --max-tokens 100 \
  --verbose

2. Using llama.cpp / GGUF v3 Runtime

./llama-cli -m qwen3.8-flash-next-q4_k_m.gguf \
  -p "Explain the AdS/CFT holographic correspondence in 2 paragraphs." \
  -n 512 --ctx-size 32768 -ngl 10

3. Using Ollama

Create a Modelfile:

FROM ./qwen3.8-flash-next-q4_k_m.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER num_ctx 32768

Register and execute:

ollama create qwen38-flash-next -f Modelfile
ollama run qwen38-flash-next "Construa um pipeline de streaming assíncrono em Rust."

---

🛡️ License & Sovereign Architecture

The weights provided are based on the Qwen architecture licensed under Apache 2.0. The ARKCOMPACT high-density ternary packing engine, hyperbolic Poincaré projections, and Triton/HIP accelerated kernels are part of the ArkheionNet Sovereign AI OS.