el4/GRM-3.2-Sky-ONYX-GGUF overview

---

base_model: OrionLLM/GRM-3.2-Sky

base_model_relation: quantized

tags:

- gguf

- ollama

- llama.cpp

- imatrix

- moe

- reasoning

- frontier

- qwen3

- onyx

pipeline_tag: text-generation

---

ollama run hf.co/el4/GRM-3.2-Sky-ONYX-GGUF:GRM-3.2-Sky-ONYX-mini.gguf

> To prove this extreme compression didn't lobotomize the model, we subjected the ~11GB ONYX-mini quant to a grueling 20-minute agentic stress test designed to break long-horizon reasoning. We tasked the model with writing a compilable Rust program to calculate quantum harmonic oscillator eigenvalues via nalgebra matrix diagonalization, forcing it to execute the code via bash, read the compiler errors, and autonomously debug its own Unicode and eigenvalue-sorting bugs. After successfully patching the math to achieve <0.05% analytical error, the model seamlessly passed the entire technical explanation through a three-step translation gauntlet into formal Japanese, Spanish, and colloquial Egyptian Arabic. The fact that a heavily crushed 35B MoE running on a consumer laptop GPU can hold the working memory required to write physics code, run cargo build, fix its own mistakes, and perfectly execute multilingual code-switching in just 20 minutes proves that ONYX doesn't just shrink the weights—it preserves the model's agentic soul.

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Geist', Roboto, sans-serif; border: 1px solid #333; border-radius: 18px; box-shadow: 0 18px 40px rgba(255, 255, 255, 0.05); overflow: hidden; background: #000000; margin-bottom: 30px;">

<div style="background: radial-gradient(circle at 15% 15%, rgba(255, 255, 255, 0.08) 0%, rgba(40, 40, 40, 0.4) 28%, transparent 46%), linear-gradient(135deg, #050505 0%, #1a1a1a 58%, #000000 100%); padding: 28px; color: white; border-bottom: 1px solid #222;">

<div style="display: flex; align-items: center; justify-content: space-between; flex-wrap: wrap; gap: 12px;">

<h1 style="margin: 0; font-size: 28px; font-weight: 900; display: flex; align-items: center; gap: 12px; color: #ffffff; border: none;"><a href="https://huggingface.co/spaces/el4/ONYX" display: flex; align-items: center; gap: 12px; style="color: #fff; text-decoration: none;">🪨 ONYX</a></h1>

<span style="background: #ffffff; color: #000000; font-size: 11px; font-weight: 900; padding: 5px 12px; border-radius: 999px; text-transform: uppercase; letter-spacing: 0.7px; box-shadow: 0 0 12px rgba(255, 255, 255, 0.2);">Adaptive Precision Engine</span>

</div>

<p style="margin: 10px 0 0 0; font-size: 14px; color: #a0a0a0; font-weight: 600;">This repo contains ONYX Quants of OrionLLM/GRM-3.2-Sky</p>

</div>

<!-- Technical Highlights -->

<div style="padding: 24px; border-bottom: 1px solid #222;">

<h3 style="margin: 0 0 16px 0; color: #fff; font-size: 16px; font-weight: 800; text-transform: uppercase; letter-spacing: 1px;">⚙️ The ONYX Architecture</h3>

<div style="display: grid; grid-template-columns: 1fr; gap: 12px;">

<div style="background: #0d0d0d; border: 1px solid #333; border-radius: 12px; padding: 16px;">

<h4 style="color: #fff; margin: 0 0 6px 0; font-size: 14px; font-weight: 700;">🧠 Dynamic Layer Sensitivity</h4>

<p style="color: #888; font-size: 12px; margin: 0; line-height: 1.5;">Replaces hardcoded edge boundaries with real activation variance measurements. ONYX autonomously identifies critical layers (like mid-network attention blocks) and protects them dynamically.</p>

</div>

<div style="background: #0d0d0d; border: 1px solid #333; border-radius: 12px; padding: 16px;">

<h4 style="color: #fff; margin: 0 0 6px 0; font-size: 14px; font-weight: 700;">🎯 Router-Weighted Imatrix</h4>

<p style="color: #888; font-size: 12px; margin: 0; line-height: 1.5;">Captures MoE router probabilities and multiplies them into activation scales. This forces the quantizer to aggressively crush "cold" experts while fiercely protecting "hot" ones within the same tensor block.</p>

</div>

<div style="background: #0d0d0d; border: 1px solid #333; border-radius: 12px; padding: 16px;">

<h4 style="color: #fff; margin: 0 0 6px 0; font-size: 14px; font-weight: 700;">🏗️ Architecture-Agnostic</h4>

<p style="color: #888; font-size: 12px; margin: 0; line-height: 1.5;">Dynamically reads HuggingFace modules and the generated F16 GGUF to map tensors. No hardcoded regex. Works out-of-the-box on Llama, DeepSeek, and custom hybrid SSM/MoE architectures.</p>

</div>

</div>

</div>

<!-- Tiers Table -->

<div style="padding: 24px;">

<table style="width: 100%; border-collapse: collapse; font-family: inherit; font-size: 13px;">

<thead>

<tr style="background: #111111;">

<th style="padding: 10px; border-bottom: 2px solid #333; text-align: left; color: #ffffff; font-weight: 900;">Tier Name</th>

<th style="padding: 10px; border-bottom: 2px solid #333; text-align: left; color: #ffffff; font-weight: 900;">Target Quality</th>

<th style="padding: 10px; border-bottom: 2px solid #333; text-align: left; color: #ffffff; font-weight: 900;">Target Size</th>

<th style="padding: 10px; border-bottom: 2px solid #333; text-align: left; color: #ffffff; font-weight: 900;">Middle Layer Strategy</th>

</tr>

</thead>

<tbody>

<tr><td style="padding: 10px; border-bottom: 1px solid #222; font-weight: 800; color: #e0e0e0;">🪨 <code style="color:#ffffff;">quality</code></td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">Q8 Match</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">~21 GB</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">IQ4_XS</td></tr>

<tr style="background: #0a0a0a;"><td style="padding: 10px; border-bottom: 1px solid #222; font-weight: 800; color: #e0e0e0;">⚖️ <code style="color:#ffffff;">balanced</code></td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">Q6 Match</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">~24 GB</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">Q5_K</td></tr>

<tr><td style="padding: 10px; border-bottom: 1px solid #222; font-weight: 800; color: #e0e0e0;">📦 <code style="color:#ffffff;">compact</code></td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">Q4 Match</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">~16 GB</td><td style="padding: 10px; border-bottom: 1px solid #222; color: #888;">Q3_K</td></tr>

<tr style="background: #0a0a0a;"><td style="padding: 10px; font-weight: 800; color: #e0e0e0;">🚀 <code style="color:#ffffff;">mini</code></td><td style="padding: 10px; color: #888;">Q2 Match</td><td style="padding: 10px; color: #888;">~12 GB</td><td style="padding: 10px; color: #888;">IQ2_XXS</td></tr>

</tbody>

</table>

</div>

</div>

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; border: 1px solid #222; border-radius: 16px; overflow: hidden; background: #000000; box-shadow: 0 10px 24px rgba(255, 255, 255, 0.03); margin-bottom: 30px;">

<div style="background: linear-gradient(135deg, #0d0d0d 0%, #1a1a1a 65%, #000000 100%); padding: 18px 20px; color: white; border-bottom: 1px solid #222;">

<h3 style="margin: 0; color: #ffffff; border: none; font-size: 19px; font-weight: 900;">📚 Credits & Foundations</h3>

</div>

<div style="padding: 20px; font-size: 13px; color: #888; line-height: 1.7;">

<p style="margin: 0 0 12px 0;">👉 <b style="color: #fff;">APEX Quantization Method</b><br>Ettore Di Giacinto & Richard Palethorpe (LocalAI Team). ONYX evolves the layer-wise precision gradients and MoE-aware tensor classification outlined in the APEX technical paper into a fully dynamic, data-driven engine.</p>

<p style="margin: 0 0 12px 0;">👉 <b style="color: #fff;">Bartowski and Lamim</b><br>For the excellent semantic imatrix calibration dataset that powers ONYX's activation scaling.</p>

<p style="margin: 0 0 12px 0;">👉 <b style="color: #fff;">llama.cpp</b><br>Georgi Gerganov and contributors for the foundational inference and quantization engine.</p>

<p style="margin: 0 0 12px 0;">👉 <b style="color: #fff;">HuggingFace Accelerate</b><br>For the <code style="color:#ddd;">init_empty_weights()</code> context manager that makes the 0-RAM "Ghost Model" possible on consumer hardware.</p>

</div>

</div>

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; border: 1px solid #444; border-radius: 18px; box-shadow: 0 18px 40px rgba(255, 255, 255, 0.04); overflow: hidden; background: #000000; margin-top: 30px; margin-bottom: 30px;">

<div style="background: radial-gradient(circle at 15% 15%, rgba(255, 255, 255, 0.1) 0%, rgba(200, 200, 200, 0.05) 28%, transparent 46%), linear-gradient(135deg, #0a0a0a 0%, #1a1a1a 58%, #333333 100%); padding: 20px; color: white;">

<h3 style="margin: 0; color: #ffffff; border: none; font-size: 18px; font-weight: 900; display: flex; align-items: center; gap: 10px;">

<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="#fff"><path d="M12 1.75L5.75 12.25L12 16L18.25 12.25L12 1.75M5.75 13.5L12 17.25L18.25 13.5L12 22.25L5.75 13.5Z"/></svg>

Support the Project

</h3>

</div>

<div style="padding: 24px; text-align: center; background: #050505;">

<p style="margin: 0 0 16px 0; font-size: 14px; color: #ccc; line-height: 1.6;">

A coffee in Ethereum would be cool! Although I don't drink coffee—I think it tastes like burnt water—but a pink lemonade would be fire! 🔥

</p>

<a href="https://etherscan.io/address/0xDEE7fa8C421BD038D32e4441ea1aDe72fE973706" target="_blank" style="display: inline-block; background: #000; border: 1px solid #666; border-radius: 12px; padding: 12px 24px; font-family: 'JetBrains Mono', monospace; font-size: 13px; color: #fff; letter-spacing: 0.5px; text-decoration: underline; text-underline-offset: 4px; transition: all 0.3s ease;">

0xDEE7fa8C421BD038D32e4441ea1aDe72fE973706

</a>

</div>

</div>

recommended sampling parameters:

--temp 0.85 --top-p 0.95 --top-k 40 --min-p 0.05 --repeat-penalty 1.1 --presence-penalty 0.0