GRM-3.2-Sky EXL3 6.08bpw HB6
This repository contains an EXL3 quantization of OrionLLM/GRM-3.2-Sky for ExLlamaV3.
Quantization
- Model: OrionLLM/GRM-3.2-Sky
- Quantizer:
exllamav3-1.4.1
- Format:
exl3
- Bits per weight:
6.08
- Head bits:
6
- Calibration:
250 rows, 2048 columns
- Output scales:
always
- Codebook:
mul1
- Files: 4 sharded
.safetensors files
Compatibility
For the original model card and non-quantized weights, see OrionLLM/GRM-3.2-Sky.
Original Model Card: GRM-3.2-Sky
<p align="center">
<img src="https://cdn-uploads.huggingface.co/production/uploads/685ea8ff7b4139b6845ce395/fLHvGnK-flK4wfEDQkgir.png" alt="logo" width="500">
</p>
<div align="center">
<a href="https://huggingface.co/OrionLLM/GRM-3.2-Sky/" style="text-decoration: none;">
<img src="https://img.shields.io/badge/🤗-HuggingFace-FC926C?style=for-the-badge" alt="HuggingFace">
</a>
<a href="https://huggingface.co/collections/OrionLLM/grm-32" style="text-decoration: none;">
<img src="https://img.shields.io/badge/📚-Collection-3B82F6?style=for-the-badge" alt="Collection">
</a>
<a href="https://grape.skinnertopia.com/chat" style="text-decoration: none;">
<img src="https://img.shields.io/badge/💬-Chat-22C55E?style=for-the-badge" alt="Chat">
</a>
<a href="https://www.apache.org/licenses/LICENSE-2.0" style="text-decoration: none;">
<img src="https://img.shields.io/badge/📜-License-E343BD?style=for-the-badge" alt="License">
</a>
</div>
1. Introduction
We're introducing GRM-3.2-Sky, our latest flagship model built for long-horizon agentic tasks and extremely difficult reasoning problems. GRM-3.2-Sky marks a substantial leap in long-horizon task capability over its predecessor, GRM-2.6-Plus, and is designed to serve as a dependable engine for complex, multi-step workflows.
The model is purpose-built for long-horizon agentic tasks and problems that are simply hard — difficult coding challenges, advanced mathematics, and rigorous logical reasoning. GRM-3.2-Sky aims to sustain coherent, goal-directed behavior over extended interactions, making it well suited for users who need a model that doesn't lose the thread across many steps of tool use, planning, and self-correction.
2. Key Capabilities
- Long-Horizon Agentic Mastery: GRM-3.2-Sky is specifically optimized to maintain coherence, planning quality, and task fidelity across long, multi-step agentic workflows, a significant step up from GRM-2.6-Plus.
- Elite Reasoning on Hard Problems: Strong performance on difficult coding, advanced mathematics, and logical reasoning tasks, with careful, structured step-by-step problem-solving.
- Robust Coding Ability: Handles complex, difficult codebases and multi-file coding tasks, including debugging, refactoring, and long-running terminal/agentic coding sessions.
- Consistent Logical Reasoning: Built to reason carefully through multi-constraint logic problems without losing track of intermediate steps.
- Flagship-Class Performance: Positioned as the top of the GRM lineup, intended to compete head-to-head with frontier-scale models on the hardest tasks.
3. Performance
GRM-3.2-Sky is designed as our most capable model to date for long-horizon agentic work and difficult reasoning. It builds directly on the strengths of GRM-2.6-Plus while specifically targeting the failure modes that emerge over long task horizons — drift, inconsistency, and loss of goal state — resulting in meaningfully improved reliability across extended sessions.
Its core strength is sustained intelligence over time: elite-level reasoning, resilient long-horizon planning, and the ability to stay on task through difficult, multi-step coding, math, and logic problems.

Detailed Benchmarks
<table>
<tr>
<th style="background: rgba(128,128,128,0.1); text-align: center;"> </th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">GRM-3.2-Sky</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">GRM-2.6-Plus</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">GPT-5.6-Luna</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">Sonnet 5</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">Gemini 3 Pro</th>
</tr>
<tr>
<td align="center" colspan="6" style="background: linear-gradient(90deg, rgba(124,58,237,0.45) 0%, rgba(99,102,241,0.42) 50%, rgba(59,130,246,0.45) 100%); font-weight: bold; height:32px; padding-top:2px; padding-bottom:2px;"><i>Knowledge & STEM</i></td>
</tr>
<tr>
<td align="center">MMLU-Pro</td>
<td align="center">89.5</td>
<td align="center">86.8</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center"><b>89.8</b></td>
</tr>
<tr>
<td align="center">MMLU-Redux</td>
<td align="center"><b>96.9</b></td>
<td align="center">94.2</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">—</td>
</tr>
<tr>
<td align="center">GPQA Diamond</td>
<td align="center">90.6</td>
<td align="center">88.3</td>
<td align="center"><b>92.3</b></td>
<td align="center">—</td>
<td align="center">91.9</td>
</tr>
<tr>
<td align="center" colspan="6" style="background: linear-gradient(90deg, rgba(124,58,237,0.45) 0%, rgba(99,102,241,0.42) 50%, rgba(59,130,246,0.45) 100%); font-weight: bold; height:32px; padding-top:2px; padding-bottom:2px;"><i>Reasoning & Coding</i></td>
</tr>
<tr>
<td align="center">LiveCodeBench v6</td>
<td align="center"><b>87.7</b></td>
<td align="center">84.8</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">82.9</td>
</tr>
<tr>
<td align="center">HMMT Feb 26</td>
<td align="center"><b>86.4</b></td>
<td align="center">84.8</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">—</td>
</tr>
<tr>
<td align="center">AIME26</td>
<td align="center"><b>96.3</b></td>
<td align="center">95.1</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">—</td>
</tr>
<tr>
<td align="center" colspan="6" style="background: linear-gradient(90deg, rgba(124,58,237,0.45) 0%, rgba(99,102,241,0.42) 50%, rgba(59,130,246,0.45) 100%); font-weight: bold; height:32px; padding-top:2px; padding-bottom:2px;"><i>General Agent</i></td>
</tr>
<tr>
<td align="center">SWE-bench Verified</td>
<td align="center">81.4</td>
<td align="center">77.7</td>
<td align="center">—</td>
<td align="center"><b>85.2</b></td>
<td align="center">76.2</td>
</tr>
<tr>
<td align="center">SWE-bench Pro</td>
<td align="center">58.3</td>
<td align="center">54.0</td>
<td align="center">62.7</td>
<td align="center"><b>63.2</b></td>
<td align="center">—</td>
</tr>
<tr>
<td align="center">Terminal-Bench 2.1</td>
<td align="center">66.3</td>
<td align="center">—</td>
<td align="center"><b>84.7</b></td>
<td align="center">80.4</td>
<td align="center">—</td>
</tr>
<tr>
<td align="center">NL2Repo</td>
<td align="center"><b>35.6</b></td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">—</td>
<td align="center">—</td>
</tr>
</table>
Scores are taken from each provider's own published model card, blog post, or system card where available; "—" indicates a score was not publicly reported by that provider at the time of writing. Different labs may use different agent scaffolds when reporting SWE-bench and Terminal-Bench results, so cross-provider comparisons should be read with that caveat.
4. Family
The GRM-3.2 family is available in various sizes to suit every use case.
<table>
<tr>
<th style="background: rgba(128,128,128,0.1); text-align: center;">Model</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">Size</th>
<th style="background: rgba(128,128,128,0.1); text-align: center;">Domain</th>
</tr>
<tr>
<td align="center"><b>GRM-3.2-Sky</b></td>
<td align="center"><b>35B-A3B</b></td>
<td align="center"><b>Flagship model for long-horizon tasks</b></td>
</tr>
<tr>
<td align="center">GRM-3.2-Cliff</td>
<td align="center">9B</td>
<td align="center">Capable model for low GPU environments</td>
</tr>
<tr>
<td align="center">GRM-3.2-Turf</td>
<td align="center">1.2B</td>
<td align="center">Lightweight model for practical reasoning</td>
</tr>
</table>
5. Architecture
GRM-3.2-Sky is built on the Ornith-1.0-35B architecture, a 35B-parameter Mixture-of-Experts model with ~3B active parameters (35B-A3B), optimized for long-horizon agentic workflows, difficult coding, advanced mathematics, and rigorous logical reasoning, while remaining efficient to deploy thanks to its sparse activation.
<div align="center">
GRM-3.2-Sky is developed by OrionLLM and released under the Apache 2.0 License.
</div>