Qwen3.8-27b-Terse-Coder-LoRA

Qwen3.8-27b-Terse-Coder-LoRA

Cuts the length of Qwen3.8-27B's reasoning on coding tasks so answers arrive sooner, as a small LoRA adapter. For people who run the base model on their own hardware.

Made by Shockem on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a small LoRA adapter for Qwen3.8-27B that makes the model think much less before it answers coding problems. It was trained with preference pairs picked by running tests and pruning steps by entropy, with no human or LLM judge. The author says it keeps correctness while the reasoning gets far shorter.

Who it suits: It suits people who run a 27B model on their own hardware and wait too long for it to finish reasoning, and researchers who study how to compress reasoning.

Industry | Who uses it | How they use it | Example
Software | Developers with a local coding model | Cut the long wait before code appears | A near-direct answer to a short coding problem
Self-hosting | Hobbyists on consumer hardware | Run the 27B base with shorter reasoning | Less time spent waiting on each prompt
Research | Machine learning researchers | Study reasoning compression with DPO on test-checked pairs | Compare round 6, 7 and 8 results

People who want to read the model's full step-by-step reasoning can skip it, since the adapter shortens that on purpose.

Where to use it:

• Python with the peft library, which loads LoRA adapters
• vLLM, which the author used with the adapter loaded at runtime
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Shorter reasoning | Developers | Answers after about 41 reasoning tokens on average in the card's test
Stock base pairing | Self-hosters | The biggest cut, on Qwen3.8-27B or its NVIDIA NVFP4 version
Signal pairing | Self-hosters | A 40 percent cut on top of an already short model

How to set it up: You need a few things first.

• Python and pip
• A Qwen3.8-27B base model, since this adapter does not work alone
• A runtime that loads LoRA adapters, such as peft or vLLM

1. Install the Hugging Face hub tool with pip.
2. Download the adapter files.
3. Load the base model, then load the adapter on top of it.

pip install -U huggingface_hub
hf download Shockem/Qwen3.8-27b-Terse-Coder-LoRA

How to use it: You prompt the base model with the adapter loaded, the same way you would prompt it without. The card ranks the pairings by strength. Stock Qwen3.8-27B or its NVIDIA NVFP4 build comes first, Signal-3.8-27B second, and a third base from the same author is optional.

Tool | What it does
Adapter at the repository root | The final round, round 8, and the one to use
archive folder | Holds earlier rounds, including round 7 and round 6
Round 6 in archive | The pass-neutral choice for one specific base the card names
Merged model | A version with the adapter already merged into the base, which the author also provides

Example: You give the adapter-loaded model a coding problem from a benchmark such as HumanEval. It answers with far fewer reasoning tokens. In the card's test that averaged about 41, against roughly 700 for the NVIDIA NVFP4 baseline. The code still passes its tests most of the time.

For newcomers and experts: A LoRA is a small add-on file that changes how a big model behaves, so you always need the base model too. For a first try, download the adapter, load it on Qwen3.8-27B and give it one coding question to see how short the reasoning is.

The adapter is rank 16 and trained with DPO. Its effect stacks on whatever conciseness the base already has. The card's test used 20 HumanEval and 20 MBPP problems held out from training, with pass at one scored by running tests, and warns that single runs swing by about 20 percent.

Base model Qwen3.8-27B loads first
LoRA adapter is applied on top
You send a coding prompt
The model reasons for far fewer tokens
Code answer comes back

Good to know:

• License: Apache 2.0.
• Cost to run: no API key or paid service named. The cost is the hardware for a 27B base model, since the adapter itself is about 58 MB at 4-bit by estimate.
• Data it reaches: only the prompts and code you give it. The card names no file, web or account access.
• Not for: work that needs long visible reasoning, or a study that continues, since the author says no further rounds are planned. Held-out pass was 70.0 percent against about 72 percent for the base.

Title: Reasoning tokens on the held-out problems, round 8
With the adapter: 41 tokens
NVIDIA NVFP4 baseline: 700 tokens

On Hugging Face, 13 November 2026

• Library: peft
• Parameters: 117M
• Weights: about 58 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 244
• Likes: 6
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Tags: peft, safetensors, gguf, lora, reasoning, coding, qwen3

Links

• Model: https://huggingface.co/Shockem/Qwen3.8-27b-Terse-Coder-LoRA
• Files: https://huggingface.co/Shockem/Qwen3.8-27b-Terse-Coder-LoRA/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Shockem/Qwen3.8-27b-Terse-Coder-LoRA

Then follow the usage in its model card. Nothing in this kit runs by itself.