Skip to content
Copyright © 2022-2026 basebox GmbH, all rights reserved.
Licensed to be used in conjunction with basebox, only.

// security

Model Register

  • Version: 1.02
  • As of: May 2026

This document lists the language models (LLMs) currently shipped by basebox and their variants. It documents origin, licence, technical modification and provisioning conditions under the EU AI Act (Regulation (EU) 2024/1689).

Provider Information

  • Company: basebox GmbH
  • Address: Bahnhofplatz 3, 86919 Utting am Ammersee, Germany
  • Contact: datenschutz@basebox.ai
  • Role under the EU AI Act: Supplier / infrastructure provider (Art. 3 No. 23 EU AI Act)

Scope of modification: Quantisation at various levels (e.g. 8-bit, 6-bit, 4-bit); no change to weights, architecture or training data.

Fine-tuning policy: basebox performs no fine-tuning or retraining. Customers may connect their own models or fine-tunes.

Models

Llama 3.3 70B

  • Developer: Meta AI
  • Version: 3.3
  • Licence: Meta Llama 3.3 Community License
  • Source URL: https://ai.meta.com/llama/
  • Modification by basebox: Quantisation (8-bit / 6-bit / 4-bit, depending on the customer's GPU resources)
  • Distribution status: Included in the delivery
  • Type: Open-source LLM – Large Language Model
  • Selection criteria: Provisioning depends on the VRAM and hardware profile of the customer server

Supported variants

  • Llama 3.3 70B Q8_0
  • Llama 3.3 70B Q6_K
  • Llama 3.3 70B Q4_K_M

Risk assessment

Minimal risk – base model without an application specification

Transparency information

  • Model card included: Yes
  • Bias statement: In accordance with the official Meta model card (v3.3)
  • Last review: 2025-10-10

OpenAI GPT 120B

  • Developer: OpenAI
  • Version: GPT 120B
  • Licence: OpenAI Commercial Terms of Service
  • Source URL: https://openai.com/
  • Modification by basebox: none
  • Distribution status: Included in the delivery
  • Type: Proprietary LLM – Large Language Model
  • Selection criteria: Provisioning depends on the VRAM and hardware profile of the customer server

Supported variants

  • OpenAI GPT 120B MXFP4

Risk assessment

Minimal risk – base model without an application specification

Transparency information

  • Model card included: Yes
  • Bias statement: In accordance with the official OpenAI model card (GPT 120B)
  • Last review: 2026-05-05

OpenAI GPT 20B

  • Developer: OpenAI
  • Version: GPT 20B
  • Licence: OpenAI Commercial Terms of Service
  • Source URL: https://openai.com/
  • Modification by basebox: none
  • Distribution status: Included in the delivery
  • Type: Proprietary LLM – Large Language Model
  • Selection criteria: Provisioning depends on the VRAM and hardware profile of the customer server

Supported variants

  • OpenAI PT 20B MXFP4

Risk assessment

Minimal risk – base model without an application specification

Transparency information

  • Model card included: Yes
  • Bias statement: In accordance with the official OpenAI model card (PT 20B)
  • Last review: 2025-05-05

Compliance notes

  • basebox ships only models that were published and licensed by the original developer.
  • The quantisation variant is selected based on the VRAM profile of the customer system.
  • basebox changes neither the training data nor the semantic capabilities of the model.
  • The customer is responsible for purpose, deployment context, data input and risk assessment under EU AI Act Art. 29.
  • Modifications, fine-tuning or retraining by the customer lead to a reassessment as a provider within the meaning of Art. 3 No. 2.

Signature block

  • Created by: basebox GmbH
  • Created on: 2025-10-15