Paraphase-2B

Paraphase-2B

A small offline model that rewrites stiff, AI-sounding English email drafts into plain prose while keeping the facts. Runs on a 4 to 6 GB device through llama.cpp.

Made by chadstthomas 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: Paraphase-2B rewrites AI-sounding English email drafts into plainer prose and keeps every fact. The model has 2.5B parameters and comes from MiniCPM5-2B through QLoRA fine-tuning. Its files are packaged as GGUF. It runs fully on your own device, offline, with no API and no upload.

Who it suits: It suits people who draft email with AI help and want it to read more naturally, on a phone, a small laptop or a small board. People with spare memory can take the larger 4B sibling instead.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Office work | Email writers | Rewrite a stiff draft into plain language | Turn a formal deadline reminder into a short note
Small business | Owners | Polish batches of messages offline | Rewrite a stack of drafts in one run
Software | Developers | Call a local rewriter from an app | Send a draft to llama-server and read the rewrite

People who write in other languages, or who work on long documents, can skip it.

Where to use it:

• llama.cpp, through llama-server
• A phone, a small laptop or a Raspberry Pi-class box
• Windows, macOS or Linux

Title: Top jobs
Use case | Who | What they get back
Email rewriting | Email writers | A plainer draft with the same facts
Batch rewriting | Small teams | Many drafts rewritten offline
Local rewrite service | Developers | A model behind a local server

How to set it up:

• Python on your computer, for the download
• The huggingface_hub package
• llama.cpp installed
• About 3.5 GB of free memory for the Q8 file, or about 2.3 GB for the Q4 file

1. Install the Hugging Face hub package.
2. Download the Q8 GGUF file, which is 2.68 GB.
3. Serve it with llama-server.
4. Send your draft with the card's rewrite prompt. No thinking flag is needed.

pip install -U huggingface_hub
hf download chadstthomas/Paraphase-2B

Title: Memory by file
File | Size | Free memory | Typical device
Q4_K_M | 1.56 GB | about 2.3 GB | A 4 GB phone or laptop, tight
Q8_0 | 2.68 GB | about 3.5 GB | A 6 GB phone or an 8 GB laptop

How to use it: Start the server with the Q8 file, then send the card's rewrite prompt with your draft. For a 4 GB device, use the Q4 file with a context of 1024.

llama-server -m paraphase-2b-Q8_0.gguf -c 2048

Example: You send a stiff reminder that expense reports are due by a set date. The model returns a shorter, plainer version with the same date and the same rule. The card shows one such rewrite and calls it a demo, not a benchmark.

For newcomers and experts: If you are new, run one short draft through the server and compare the before and after. Check that every number and date survived.

For experts, the card compares the 2B and 4B files on 50 paired items. The 2B with retry gave 48 valid outputs against 50, with fewer AI tells at 6.3% against 8.0%. Greedy decoding alone was only 82% valid, and the retry wrapper fixed 7 of the 9 bad rows.

Download the Q8 GGUF file
Serve it with llama-server
Send a draft with the rewrite prompt
Check facts, retry if invalid

Good to know:

• Retry needed: the 2B reaches its best validity only with the retry wrapper.
• Detector tests: the card calls its AI-detector result suggestive, not confirmed, and makes no claim of guaranteed evasion.
• License: Apache 2.0.
• Cost to run: none named, and everything stays local.
• Data it reaches: it downloads files from Hugging Face, and the card says drafts are not uploaded.
• Not for: non-English text, long documents, legal or medical advice, or sending email unattended.

Title: Valid drafts out of 50
2B Q8 greedy: 41 drafts
2B Q8 with retry: 48 drafts
4B Q8 with retry: 50 drafts

On Hugging Face, 13 November 2026

• Task: text generation
• Library: llama.cpp
• Parameters: 2.5B
• Weights: about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 135
• Likes: 4
• License: apache-2.0
• Languages: en
• Based on: openbmb/MiniCPM5-2B
• Tags: llama.cpp, safetensors, gguf, humanizer, ai-humanizer, text-humanization, text-rewriting, rewriting, paraphrase, paraphraser, style-transfer, email

Links

• Model: https://huggingface.co/chadstthomas/Paraphase-2B
• Files: https://huggingface.co/chadstthomas/Paraphase-2B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download chadstthomas/Paraphase-2B

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