LionBox

Docs · Model & deployment card

The LionBox model, and how to use it well.

What the AI model on LionBox is for, where it runs, what it gets wrong and how your team should check its work.

At a glance

Qwen 3.5 · LionBox build

Model
LionBox build of Qwen 3.5, optimized and tuned in-house for the LionBox hardware
Base model
Qwen 3.5 by Alibaba Cloud (Qwen team)
Runs on
The LionBox graphics card, inside your office
Inputs
Text and images
Outputs
Text
Version
To be confirmed before dispatch

Model

LionBox runs our own build of Qwen 3.5, optimized and tuned in-house for the LionBox hardware. The base model is developed by Alibaba Cloud (Qwen team) and is provided under its own license; see Open-source & third-party licenses.

Where it runs

The model runs entirely on the box, on its own graphics card. It does not use external AI services or inference APIs, and it keeps working with the internet disconnected. You can verify offline operation yourself.

Inputs and outputs

The model accepts text and images and answers in text. Employees can add PDF, DOCX and text files to chats and knowledge spaces.

Good for

  • Everyday chat and questions at work.
  • Answering questions about company documents, with the sources cited.
  • Reading scanned documents (OCR) and extracting information from them.
  • Drafting emails and replies.
  • Summarising meetings from recordings and transcripts.
  • Translating business material.
  • Explaining tables, figures and other business data.

Very complex, multi-step tasks, writing complex software and in-depth research. A model compact enough to run on one office appliance does not deliver results at the level of the largest cloud models on these tasks. Use it to draft, summarise and look things up, and give demanding work to a specialist or a dedicated tool.

Known limitations

  • It can produce answers that are wrong, incomplete or made up (hallucinations), and still sound convincing.
  • Its general knowledge stops at a cut-off date and does not include recent events. Internet search is off by default.
  • Answer quality can vary between languages.

Human oversight

Check answers before acting on them, especially for legal, financial or medical content and for anything sent outside your company. Review the cited source and its revision date before acting. Validate totals, dates and identifiers against the original document. Confirm commitments and recipients before sending a drafted email. Use qualified review for legal documents, regulated material and publication, and recalculate figures with appropriate tools.

Intended use

LionBox is a general-purpose office assistant. It is not designed or intended for high-risk uses listed in Annex III of the EU AI Act, such as the recruitment, selection or evaluation of workers, the assessment of the creditworthiness of natural persons, or medical diagnosis. Do not use it to infer or recognise the emotions of employees or other people, or to monitor the individual behaviour or performance of employees. See Terms of sale, section 09.

Performance

Measured on LionBox V1 with more than 2,000 benchmark messages from simulated colleagues: no request failed, the first words typically appeared in 0.07 seconds, and when 16 people pressed send at the same moment each of them saw the first words in under half a second. LionBox is recommended for teams of 1-30 people.

Compared with the public release of the model, our build gives almost twice as much room for conversations and documents and answers about 20% faster, with no measurable loss in quality across 800 reviewed answers. Results depend on the workload and how many people use the box at once.

Data handling

How prompts, answers and documents are kept on the box is described on the Security page.

Back to the docs