---

license: apache-2.0

base_model: Qwen/Qwen3.8-Flash-Next

library_name: gguf

tags:

• text-generation
• gguf
• quantizer
• autoround
• architecture-aware
• mixture-of-experts
• moe
• mamba
• ssm
• reap
• expert-pruning
• 113b

---

<!-- zerodigest-lab-banner-v2-mid-tone -->

<div style="background: #1f1f23; color: #f4f4f5; padding: 20px; border: 1px solid #38bdf8; border-radius: 8px; margin: 20px 0; font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;">

<h2 style="color: #38bdf8; margin: 0 0 10px 0; font-size: 18px; font-weight: 700; letter-spacing: -0.5px;">⚡ Fuel the Lab: Keep the Optimization Loops Running</h2>

<p style="font-size: 14px; line-height: 1.5; margin: 0 0 16px 0; color: #d4d4d8;">Every single <b>ZeroDigest YMQ</b> release is handcrafted and manually calibrated via intensive importance-matrix sweeps to protect critical logic pathways. This project is entirely independent research—no automation bots, no corporate backers, and no external funding. Running multi-hour compute arrays consumes massive local infrastructure overhead out-of-pocket. Consider checking out our compiler or supporting our compute costs!</p>

<div style="background: #18181b; padding: 12px; border-radius: 6px; border: 1px solid #3f3f46; font-size: 13px;">

<p style="margin: 0 0 8px 0; font-weight: 600; color: #f4f4f5;">👉 Developer Resources & Support Paths:</p>

<div style="margin-bottom: 6px;">🛠️ <a href="https://github.com/minyor/ymq-compiler" target="_blank" style="color: #38bdf8; text-decoration: underline; font-weight: bold;">View Compiler Source on GitHub (YMQ v2.0)</a></div>

<div style="margin-bottom: 6px;">🚀 <a href="https://runpod.io?ref=dbjmkmeh" target="_blank" style="color: #38bdf8; text-decoration: underline; font-weight: bold;">Deploy weights on RunPod Cloud Compute (Affiliate Link)</a></div>

<div style="margin-bottom: 8px;">☕ <a href="https://ko-fi.com/zerodigest" target="_blank" style="color: #38bdf8; text-decoration: underline; font-weight: bold;">Support the Lab on Ko-fi (One-Time / Monthly)</a></div>

<div style="margin: 8px 0; border-top: 1px solid #3f3f46;"></div>

<div style="margin-bottom: 4px; font-family: monospace;">🧡 <span style="color: #f97316; font-weight: bold;">BTC:</span> <code style="background: #27272a; padding: 1px 4px; border-radius: 4px; color: #f4f4f5;">14Fmic9z3VA1ZU11bWoP6JtU7csTAApwZo</code></div>

<div style="font-family: monospace;">🔷 <span style="color: #6366f1; font-weight: bold;">ETH/USDT:</span> <code style="background: #27272a; padding: 1px 4px; border-radius: 4px; color: #f4f4f5;">0xbe4cdc3adc27c21ef1c27c6b403311db07b35ed2</code></div>

</div>

</div>

Qwen3.8-Flash-Next-REAP-320-YMQ-GGUF

> Base Model: Qwen/Qwen3.8-Flash-Next

> REAP Manifest: AnonimousA/seleccion_mass_K320.json

> Inspired By: AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF

🎯 REAP Expert Pruning × Architecture-Aware Mixed Precision Quantization

This repository features a custom expert-pruned and architecture-aware quantized build of Qwen3.8-Flash-Next, processed through a two-stage pipeline:

1. REAP Expert Selection (K=320): A custom importance-matrix-driven manifest (seleccion_mass_K320.json) surgically selects the top 320 expert slots per MoE routing layer from the original dense expert pool, permanently pruning the least-activated pathways. This is not a uniform prune—selections are driven by per-token activation mass collected across calibration prompts.
2. YMQ-Compiler (v2.0) Log-Space Quantization: The pruned checkpoint is then quantized using the custom YMQ-Compiler framework, which assigns per-layer precision tiers based on importance-matrix cluster analysis—elevating high-leverage expert layers to Q6_K / IQ4_NL shields while compressing low-activity background experts down to IQ2_XS.

These builds natively support parallel multi-token prediction (MTP) speculation engines and utilize high-context optimization parameters tailored for demanding code development API execution environments (such as RooCode/Aider).

<p align="center">

<img src="./logo.jpg" width="800" alt="ZeroDigest YMQ Logo">

</p>

---

📊 Quantization Preset Tier Details

| Preset Tier | Total Size | Target Usage / Memory VRAM Profile | Cognitive Real-World Coding Quality |

| :--- | :--- | :--- | :--- |

| XS-Pro |~63 GB | 💎 48GB+ Workstation Driver | The REAP MoE High-Context. Q8_0 shared-expert armor and Q5_K attention shields protect every routing decision. The surviving expert pool is graded across four precision tiers — individually tuned per layer via importance-matrix log-space clustering. No uniform blanket compression. |

| M-TI | ~63 GB | 💎 48GB+ Workstation Driver (Recommended) | The REAP MoE High-Context. Q8_0 shared-expert armor and Q5_K attention shields protect every routing decision. The surviving expert pool is graded across four precision tiers — Q6_K ➔ IQ4_NL ➔ IQ3_XXS ➔ IQ2_XS (floor) — individually tuned per layer via importance-matrix log-space clustering. No uniform blanket compression. |

📉 Perplexity Evaluation Metrics (WikiText-2)

The following metrics demonstrate the mathematical quality preservation of the REAP + YMQ-Compiler pipeline compared to standard linear quantization layouts. Tested natively via llama-perplexity over a 4096 context window using the official WikiText-2 test corpus.

| Model Variant | File Size | Perplexity | Mean KL-Divergence | Internal Bit Gradient (High ➔ Mid ➔ Low ➔ Floor) |

| :--- | :--- | :--- | :--- | :--- |

| XS-Pro | ~63 GB | TBD | TBD | Expert Gradient: IQ4_XS ➔ IQ3_S ➔ IQ2_S ➔ IQ2_XS (floor) · Armor: Q8_0 / Q5_K |

| M-TI | ~63 GB | TBD | TBD | Expert Gradient: Q6_K ➔ IQ4_NL ➔ IQ3_XXS ➔ IQ2_XS (floor) · Armor: Q8_0 / Q5_K |

> ℹ️ Why so large? Qwen3.8-Flash-Next is a massive sparse Mixture-of-Experts architecture with hybrid Mamba SSM state-space tracking. Even after REAP expert pruning (K=320 per layer) and aggressive 2-bit expert floors, the surviving expert pool keeps the file at ~63 GB. The active parameter count per token remains small (~5B), so inference speed stays comparable to much smaller dense models once weights are resident in memory.

> ⚠️ REAP Pruning Trade-Off: REAP expert pruning is not free. Permanently removing ~40% of the least-activated expert slots means the knowledge those experts carried is gone forever—it cannot be recovered by quantization or fine-tuning. Expect noticeable degradation on broad general-knowledge tasks, rare factual recall, and domain-specific queries that relied on the pruned pathways. The K=320 selection threshold keeps the experts that fire most often across calibration prompts, but "rarely activated" is not the same as "useless." For coding, structured reasoning, and agent-loop workloads—where the surviving expert pathways carry the bulk of the signal—this build remains a strong and practical driver. If your primary use case is code generation, refactoring, or long-context agent execution, the pruned model handles those tasks reliably. For open-ended creative writing or encyclopedic recall, expect the full unpruned model to perform better.

💡 The REAP + YMQ Pipeline Breakthrough

Standard quantization pipelines treat every expert in a Mixture-of-Experts model as equally important, applying a uniform bit-depth across the entire network. This wastes valuable VRAM on dormant background experts that rarely activate, while starving the critical routing pathways that actually carry reasoning signal.

REAP (Router-based Expert Activation Pruning) solves this at the architectural level:

• Expert Mass Profiling: During calibration, every expert slot in every MoE layer is scored by cumulative activation mass across representative prompts. The seleccion_mass_K320.json manifest captures the top-320 highest-mass expert indices per layer.
• Permanent Pruning: Expert slots that fall below the selection threshold are permanently removed from the checkpoint, reducing the model's effective parameter count without any distillation or retraining loss.
• Importance-Aware Quantization: The same importance data that drove pruning is reused by the YMQ-Compiler to assign per-layer precision tiers across a four-grade gradient: high-mass expert layers receive full Q6_K armor, mid-tier layers are elevated to IQ4_NL, lower-importance blocks drop to IQ3_XXS, and dormant background layers are safely compressed to the IQ2_XS floor.

The result is a model that retains the reasoning capacity of the full 113B-parameter dense MoE within a ~63 GB footprint, with per-layer quantization budgets precisely matched to actual cognitive load.

---

⚖️ YMQ vs. Uniform Quantization (The AutoRound Philosophy)

Standard quantization pipelines apply a blunt, uniform bit-depth across every single layer in a model. This wastes valuable VRAM on silent background layers while starving critical logic anchors of necessary precision.

The YMQ-Compiler implements a philosophy similar to advanced weight-tuning frameworks like Intel's AutoRound:

• Targeted Bit Allocation: It strips bits away from low-leverage background tensors and automatically re-allocates that saved VRAM budget straight into full high-fidelity shields for the model's highest cognitive spikes and boundary pathways.
• Instant Optimization: Instead of running heavy, days-long optimization training loops, YMQ achieves a highly accurate mixed-precision layout instantly by analyzing layer importance metrics in log-space.

The result is a custom mixed-precision portfolio that matches the low perplexity and high context stability of premium optimized quants (like AutoRound), while maintaining an ultra-lightweight, high-speed cache footprint.

---

🛠️ The YMQ Compilation Architecture

Standard quantization pipelines treat network tensors like a flat dataset, applying destructive blanket low-bit compression to delicate tracking networks. The YMQ-Compiler solves high-context logic decay by parsing model files dynamically via an automated, multi-tiered protection matrix:

1. REAP-Aware Per-Layer Expert Grading: Unlike standard quantizers that apply one bit-depth to all ffn_*_exps tensors, the YMQ-Compiler reads the importance manifest and assigns each of the 48 transformer blocks a unique expert precision tier across a four-grade gradient: HIGH (Q6_K) → MID (IQ4_NL) → LOW (IQ3_XXS) → FLOOR (IQ2_XS). High-mass layers like blk.0, blk.4, blk.6, and blk.47 receive full Q6_K armor, while dormant background blocks safely compress to the IQ2_XS floor.
2. Log-Space Gap Detection Clustering: Instead of flat percentage thresholds, the engine computes statistical cluster variances in log-space, isolating intermediate logical reasoning spikes and elevating them to stable non-linear formats, while compressing idle fact-storage expert layers to aggressive 2-bit baselines.
3. Fading Boundary Tapering: Recognizes the extreme fragility of initial token entry data vectors, forcing an input wave cushion that gradually stabilizes parameters before hitting the fallback pools.
4. Dedicated Gate Insulation: Hard-shields volatile parallel Transformer Multi-Head Attention, Mamba Linear State Space Model (SSM), Indexer, and PLE (Positional Layer Embedding) routing paths at Q8_0/Q5_K, keeping context tracking perfectly noise-free across the hybrid architecture.
5. Shared Expert Armor: The always-active shared expert pathways (ffn_gate_shexp, ffn_up_shexp, ffn_down_shexp) are locked at Q8_0—effectively lossless—to preserve the reasoning backbone that every token traverses regardless of expert routing decisions.
6. Asymmetric Vocabulary Shielding: Fixes tied-weight boundary errors by mapping the final logit classification exit heads to robust configurations to completely eliminate formatting loops and API tag leakage under deep contexts.
7. Native Next-N Speculative Stripping: Processed with advanced pre-tokenizer stripping to ensure zero index offset drift or layer-shifting risks across hybrid REAP-pruned configurations.

---

🌾 The REAP Pruning Pipeline

This build was produced by a custom two-stage pipeline:

| Stage | Tool | Description |

| :--- | :--- | :--- |

| 1 — Expert Pruning | ymq_reap_gguf.py | Reads the seleccion_mass_K320.json manifest and permanently prunes all non-selected expert slots from the base BF16 checkpoint. Reduces expert pool density while preserving activation-mass pathways. |

| 2 — Importance-Aware Quantization | ymq-compiler | Applies per-layer precision tiers derived from importance-matrix log-space clustering. Shared experts and attention gates receive high-fidelity shields; background expert pool is compressed to dense 2-bit grids. |

> REAP Manifest Source: The expert selection manifest (seleccion_mass_K320.json) was authored by AnonimousA as part of the Qwen3.8-Flash-Next-REAP-320-GGUF project. Credit and thanks for the calibration methodology.

---

🚀 Recommended Runtime Parameters (llama.cpp / llama-server)

Need to scale up? Deploy this exact script on on-demand cloud GPUs via RunPod Cloud Compute.

$ ./llama-server -m Qwen3.8-Flash-Next-reap320-YMQ-M-TI.gguf \
  --ctx-size 145000 \
  --ctx-checkpoints 2 \
  -ctk q8_0 -ctv q4_0 \
  --spec-type draft-mtp --spec-draft-n-max 2 \
  --md mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf \
  --mmproj Qwen3.8-Flash-Next-mmproj-Q4_K_S.gguf \
  --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 \
  --jinja -fa \
  --ubatch-size 512 \
  --chat-template-kwargs '{"reasoning_effort": "low"}'

> ⚠️ VRAM Note: At ~63 GB, this model runs comfortably on 48GB of vram or x2 3090.

🖼️ Vision Projection (--mmproj)

For multimodal vision support, pair this build with the native Flash-Next vision projection:

| Variant | File | Size | Notes |

| :--- | :--- | :--- | :--- |

| Q4_K_S | Qwen3.8-Flash-Next-mmproj-Q4_K_S.gguf | ~478 MB | Compact vision projection for VRAM-constrained setups. Retains strong image understanding at reduced footprint. |

Pass via --mmproj <path-to-file> in your llama-server invocation (see example above).

🎕 MTP Speculative Draft (--md)

This model ships with native multi-token prediction heads. Pair with the shared MTP draft head for speculative decoding:

| Variant | File | Size | Notes |

| :--- | :--- | :--- | :--- |

| Q4_K_M | mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf | ~2 GB | Shared MTP draft head for 2-token speculative decoding. Enable via --spec-type draft-mtp --spec-draft-n-max 2. |

---

☕ Support & Future R&D

If the YMQ-Compiler builds saved your context window from collapsing or optimized your active development cycle speeds, consider buying a coffee to fund further low-level optimization research. Your support keeps the server nodes baking future model scales!

👉 Support ZeroDigest Research on ko-fi

---

📦 Source Framework & Automation Code

The compiler pipeline automation engine, REAP pruning scripts, setup thresholds, and structural mapping rules are open-source. To view the implementation details or compile your own custom models natively using this profile layout, visit the official development hub:

👉 GitHub: ZeroDigest / YMQ-Compiler

---

🙋 Acknowledgements

| Credit | Role |

| :--- | :--- |

| Qwen/Qwen3.8-Flash-Next | Base architecture and weights |

| AnonimousA | REAP expert selection manifest (seleccion_mass_K320.json) and calibration methodology inspiration |

| YMQ-Compiler | Architecture-aware importance-matrix quantization pipeline (this build) |