zephyr-7b-beta

zephyr-7b-beta

Chats as a helpful assistant, built by tuning Mistral-7B with direct preference optimization. For developers and hobbyists who want a small open chat model they can run themselves.

Made by HuggingFaceH4 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: Zephyr-7B-beta is the second model in HuggingFaceH4's Zephyr series of chat assistants. It is a 7.2 billion parameter fine-tune of Mistral-7B-v0.1, trained on public synthetic datasets with direct preference optimization (DPO). The card says removing the built-in alignment from those datasets helped its scores and made it more helpful.

Who it suits: It suits people who want a compact open chat model, from developers adding a self-hosted assistant to researchers studying how DPO shapes a model. The record lists English.

Industry | Who uses it | How they use it | Example
Software | App developers | Add a small chat assistant they host themselves | A support bot with a custom persona
Research | Alignment researchers | Study a model tuned with DPO on public synthetic data | Compare it with its alpha predecessor
Hobby | Local model users | Chat with a 7B model on their own machine | Ask everyday questions in a chosen style

People who need strong coding or maths help can skip it, since the card says it trails proprietary models on those tasks.

Where to use it:

• Python with the Hugging Face transformers library, on Windows, macOS or Linux
• A hosted demo chat that the model card points to
• Your own machine, since the weights are about 14.5 GB in BF16 or an estimated 3.6 GB at 4-bit

Use case | Who | What they get back
General chat | Hobbyists | Helpful answers in conversation
Persona bots | Developers | Replies in a style set by a system message
Preference tuning research | Researchers | A reference model trained with DPO on UltraFeedback

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

• Python and pip
• The transformers library and torch
• Memory or disk for the weights, about 14.5 GB in BF16

1. Install the Hugging Face hub tool with pip.
2. Download the model files to your computer.
3. Load the model in Python with the transformers pipeline, as the card shows.

pip install -U huggingface_hub
hf download HuggingFaceH4/zephyr-7b-beta

How to use it: You build a pipeline for text generation, then pass a list of messages. A system message sets the persona and a user message holds the question. The card formats them with the tokenizer's chat template before generating.

import torch
from transformers import pipeline

pipe = pipeline("text-generation", model="HuggingFaceH4/zephyr-7b-beta", torch_dtype=torch.bfloat16, device_map="auto")

messages = [
    {
        "role": "system",
        "content": "You are a friendly chatbot who always responds in the style of a pirate",
    },
    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]

Tool | What it does
pipeline() | Loads the model for text generation in transformers
torch_dtype=torch.bfloat16 | Loads the weights in BF16
device_map="auto" | Lets the library place the model on available devices
messages list | Holds the system persona and the user question
Chat template | Formats the messages the way the model expects

Example: You set the system message to a friendly chatbot that answers like a pirate, then ask how many helicopters a human can eat in one sitting. The model returns a text reply, which you should expect in that pirate style.

For newcomers and experts: If you are new, a chat model is a program that answers typed questions. Download the files, run the pipeline example above, and change the system message to see how the reply style changes.

Underneath, the model was first fine-tuned on a filtered version of UltraChat, a set of synthetic dialogues made by ChatGPT. It was then aligned with the TRL DPOTrainer on UltraFeedback, which holds 64k prompts with completions ranked by GPT-4. The card reports an MT-Bench score of 7.34 and an AlpacaEval win rate of 90.60 percent.

Mistral-7B-v0.1 base model
Fine-tune on filtered UltraChat dialogues
DPO on UltraFeedback rankings
Chat through the tokenizer template

Good to know:

• License: MIT.
• Cost to run: no API key or paid service named. You supply the hardware, and the weights download from Hugging Face.
• Data it reaches: only the prompts you pass in. The card names no file, web or account access.
• Not for: complex coding and maths, where the card says it trails proprietary models. The card also warns it is likely to produce problematic text when asked to.

Title: MT-Bench score at release
Mistral-Instructv0.1: 6.84 points
Zephyr-7b-alpha: 6.88 points
Llama2-Chat 70B: 6.86 points
Zephyr-7b-beta: 7.34 points
GPT-3.5-turbo: 7.94 points
GPT-4: 8.99 points

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 7.2B
• Weights: about 14.5 GB in BF16; about 3.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 60,487
• Likes: 1,868
• License: mit
• Languages: en
• Based on: mistralai/Mistral-7B-v0.1
• Quantized versions: 37
• Fine-tunes: 165
• Tags: transformers, pytorch, safetensors, mistral, text-generation, generated_from_trainer, conversational, model-index

Links

• Model: https://huggingface.co/HuggingFaceH4/zephyr-7b-beta
• Files: https://huggingface.co/HuggingFaceH4/zephyr-7b-beta/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download HuggingFaceH4/zephyr-7b-beta

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