CoolFace
Modelpublic

OrionLLM/LRM-3.2

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
18likes169downloads
Model Card

1. Introduction

We're introducing LRM-3.2, a reasoning model built around a single idea: thinking should be fast, direct, and dense — not padded. LRM-3.2 keeps the full depth of chain-of-thought reasoning while stripping out the narrative scaffolding that most models use to fill space.

Same reasoning. Same depth. Way fewer tokens. LRM-3.2 throws the grammar padding in the fire and keeps all the brain meat. The final answer still comes out in normal, full-quality English — the compressed voice lives only inside the thinking process.

2. Key Capabilities

  • —Adaptive Depth: Thinking length scales with task difficulty, not with habit. Easy problems get a one-line think; hard problems still get full, structured deliberation.
  • —Unchanged Output Quality: Compression happens exclusively in the reasoning trace. Final answers remain complete, natural, and equivalent in quality to verbose-thinking models.
  • —Fast, Direct Inference: Dramatically shorter think blocks translate directly into lower latency and lower token spend per response, without a distillation-style drop in capability.
  • —Reliable Under Long Sessions: Reasoning stays dense and on-task across extended agentic and multi-step work, rather than drifting into repetitive verbal habits.

3. Performance

<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;">LRM-3.2</th> <th style="background: rgba(128,128,128,0.1); text-align: center;">Qwen3.6-27B</th> </tr> <tr> <td align="center" colspan="3" 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 &amp; Coding</i></td> </tr> <tr> <td align="center">GSM8K</td> <td align="center"><b>95.8</b></td> <td align="center">—</td> </tr> <tr> <td align="center">HumanEval</td> <td align="center"><b>86.9</b></td> <td align="center">—</td> </tr> </table>

"—" indicates a score not yet measured on this harness. Both benchmarks reflect matched-quality answers between LRM-3.2 and its base model — the difference lives in the think trace, not the final result.

4. Efficiency in Practice

Same problem, same correct solution, radically different think length.

Task: write separate_paren_groups, a function that splits a string of parentheses into its top-level balanced groups.

Qwen3.6-27B think: 6,539 tokens. Starts like this and keeps going for pages:

<div style="background: rgba(124,58,237,0.10); border: 1px solid rgba(124,58,237,0.35); border-left: 4px solid #7C3AED; border-radius: 10px; padding: 12px 16px; margin: 8px 0;"> The user wants a Python function <code>separateparengroups</code> that takes a string of parentheses and spaces, and returns a list of strings. Each string in the list should represent a balanced group of parentheses that is not nested within another group. Spaces should be ignored... </div>

LRM-3.2 think: 33 tokens. The whole thing:

<div style="background: rgba(124,58,237,0.10); border: 1px solid rgba(124,58,237,0.35); border-left: 4px solid #7C3AED; border-radius: 10px; padding: 12px 16px; margin: 8px 0;"> Strip spaces. Scan chars; depth counts open parens. When depth becomes 0 after a close, current group finished; append and reset. Empty input -> []. </div>

Same answer quality. 198x less think.

5. Training

LRM-3.2 is fine-tuned from Qwen3.6-27B on the grug-think and grug-think-v3-10k datasets, applying a think-only loss on trajectory data so the compressed reasoning style is learned without touching final-answer quality.

LRM-3.2 is directly inspired by ProCreations/grug-27b, which pioneered this padding-free thinking approach on the same base model family. LRM-3.2 adapts that approach under the OrionLLM naming and evaluation pipeline.

6. Architecture

LRM-3.2 is built on Qwen3.6-27B, a 27B-parameter dense model, fine-tuned to compress the reasoning trace while leaving final-answer generation untouched. No changes are made to the base tokenizer, context length, or output formatting — only the internal thinking style is altered.

<div align="center">

Orion Research - 2026

</div>