What are the hardware requirements for GEITje 7B Ultra?

GEITje 7B Ultra requires approximately 14.4 GB of VRAM and supports a context window of 32K tokens. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed GEITje 7B Ultra and how large is it?

GEITje 7B Ultra is developed by BramVanroy, a model with 7b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

How does GEITje 7B Ultra perform on standard benchmarks?

GEITje 7B Ultra has the following published scores: MMLU 50.16. Compare against reference models on this page or on the LLM Explorer leaderboards.

GEITje 7B Ultra by BramVanroy — VRAM 14.4GB, 32K context

Name: GEITje 7B Ultra
Author: BramVanroy

GEITje 7B Ultra is an open-source language model by BramVanroy. Features: 7b LLM, VRAM: 14.4GB, Context: 32K, License: cc-by-nc-4.0, LLM Explorer Score: 0.18, Arc: 45.5, HellaSwag: 75.5, MMLU: 50.2, GSM8K: 19.4.

Arxiv:2412.04092 Alignment-handbook Base model:bramvanroy/geitje-7... Base model:finetune:bramvanroy... Conversational Dataset:bramvanroy/ultra feedb... Deploy:azure Dpo Geitje Mistral Nl Region:us Safetensors Sharded Tensorflow Trl

Model Card on HF 🤗: https://huggingface.co/BramVanroy/GEITje-7B-ultra

GEITje 7B Ultra Parameters and Internals

Model Type

text-generation, conversational

Use Cases

Areas:

research, non-commercial applications

Applications:

chatbot, conversational interface

Primary Use Cases:

Dutch conversational model

Limitations:

Model might generate inaccuracies and potentially offensive content

Supported Languages

nl (Dutch)

Training Details

Data Sources:

BramVanroy/ultra_feedback_dutch

Data Volume:

56,137,090 tokens

Methodology:

Based on Mistral 7B, further pretrained on Dutch data

Context Length:

2048

Training Time:

approximately 11 hours

Hardware Used:

Two nodes of four A100 80GB each

Model Architecture:

Based on Mistral 7B and fine-tuned using DPO

Input Output

Input Format:

Conversations with roles (system, user)

Accepted Modalities:

text

Output Format:

Textual responses

Performance Tips:

Use `load_in_8bit` for reduced memory usage on GPUs; ensure compatibility with flash attention 2 if possible.

LLM Name	GEITje 7B Ultra
Repository 🤗	https://huggingface.co/BramVanroy/GEITje-7B-ultra
Base Model(s)	GEITje 7B Ultra Sft BramVanroy/GEITje-7B-ultra-sft
Model Size	7b
Required VRAM	14.4 GB
Updated	2026-08-25
Maintainer	BramVanroy
Model Type	mistral
Model Files	4.9 GB: 1-of-3 5.0 GB: 2-of-3 4.5 GB: 3-of-3 0.0 GB
Supported Languages	nl
Model Architecture	MistralForCausalLM
License	cc-by-nc-4.0
Context Length	32768
Model Max Length	32768
Transformers Version	4.36.2
Tokenizer Class	LlamaTokenizer
Padding Token	</s>
Vocabulary Size	32000
Torch Data Type	bfloat16

Rank the GEITje 7B Ultra capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Best Alternatives to GEITje 7B Ultra

Best Alternatives	Context / RAM	Downloads	Likes
...Nemo Instruct 2407 Abliterated	1000K / 24.5 GB	220	20
MegaBeam Mistral 7B 512K	512K / 14.4 GB	8470	54
SpydazWeb AI HumanAI RP	512K / 14.4 GB	16	1
SpydazWeb AI HumanAI 002	512K / 14.4 GB	18	1
...daz Web AI ChatML 512K Project	512K / 14.5 GB	12	0
MegaBeam Mistral 7B 300K	282K / 14.4 GB	3779	16
MegaBeam Mistral 7B 300K	282K / 14.4 GB	8018	17
Hebrew Mistral 7B 200K	256K / 30 GB	23	15
Astral 256K 7B	250K / 14.4 GB	9	0
Astral 256K 7B V2	250K / 14.4 GB	8	0

Note: green Score (e.g. "73.2") means that the model is better than BramVanroy/GEITje-7B-ultra.