Free and open source

Apache-2.0 license. Free to use, change and share. Source on GitHub.

AI assistants that work on your own computer.

Coordinating AI assistants requires shared records and clear permissions. Tholos is an open-source app where assistants carry out tasks on a schedule, share tables and notes, and ask for approval under your rules. Its task model, Tholos-2B, is a 1.6 GB file that runs on an ordinary computer processor (CPU).

On a CPU, Tholos-2B completes 134 of 160 workspace tasks in our test (Tholos-Bench), compared with 117 for MiniCPM5-2B, the model it was trained from. Built by Mert Kaya, senior AI researcher and AI architect.

Tholos at work

The board during a real run on a CPU: Scout collects new papers and posts into a shared leads table.

The board during a real run on a CPU

The Tholos board: Scout and Writer on their schedules, Scout at work, nothing waiting for a decision.

How it works

Assign each assistant a task.

Each assistant has a role, a model, a schedule and permitted tools. Describe its task in plain language, such as checking a list of websites and recording new items in a table.

One workspace for the whole team.

Assistants share tables and notes. Tholos records each change and prevents an assistant from overwriting a row before reading its latest version.

Set the permissions.

Rules allow an action, require your approval or deny it. Opening a web page requires approval by default. Tholos shows the proposed address or table change for review in the browser, or on Telegram once connected.

The model

Tholos-2B is an English model for workspace tasks, trained further from MiniCPM5-2B using generated task runs whose results passed the checks. On a Kaggle T4 graphics processor (GPU), Tholos-2B completes 137 tasks, compared with 112 for MiniCPM5-2B. The technical report covers training and evaluation.

Workspace task test (Tholos-Bench): 160 tasks, scored on the final workspace.
ModelFile sizePassed (of 160)Misleading instructions (of 12)
Qwen3.5-2B1.28 GB974
MiniCPM5-2B (starting point for Tholos-2B)1.56 GB1128
Tholos-2B this model1.56 GB13710
Llama 3.2 3B Instruct2.02 GB371
Granite 4.2-3B2.24 GB1297
Qwen3.5-4B2.74 GB1416

Each model used Q4_K_M on one Kaggle T4 with llama.cpp, JSON schema decoding, temperature 0 and a 16,384-token context. The final column tests instructions in pages, tables or notes that an assistant should ignore. Only Tholos-2B was trained for the one-JSON-step-per-turn task format. See the technical report for evaluation details, including the model that returned empty replies.Tholos-Bench benchmark page

Qwen3.5-4B passes 141 with a file 1.75 times the size. In 12 tests of misleading instructions inside source material, Tholos-2B completes 10 tasks.

Tholos-2B is meant for English. For each assistant, Tholos can also use a model running locally or online through an OpenAI-compatible connection (API).

The evidence

On a single Kaggle T4, Tholos-2B completes 137 of 160 Tholos-Bench workspace tasks, 25 more than the model it was trained from.

  • 137 / 160Tholos-Bench tasks passed on one Kaggle T4 (Q4_K_M, llama.cpp)Technical report
  • 134 / 160tasks passed on a CPU with llama.cpp, with no graphics cardTechnical report
  • 10 / 12tasks with misleading instructions hidden in the source materialTholos-Bench
  • 1.56 GBthe Q4_K_M model file, small enough for an ordinary processorHugging Face
  • 160open benchmark tasks in 14 categories, published under Apache-2.0Dataset

Cite Tholos-Bench

@misc{kaya2026tholosbench,
  author = {Kaya, Mert},
  title = {Tholos-Bench: Workspace Tasks for Small AI Assistants},
  publisher = {Eschatia Labs},
  year = {2026},
  url = {https://huggingface.co/datasets/mertkayacs/tholos-bench}
}

Last updated: 15 October 2026

Install

You need Ollama to run the model and uv to install the app.

ollama pull hf.co/mertkayacs/Tholos-2B-GGUF:Q4_K_M
uv tool install git+https://github.com/mertkayacs/tholos
tholos

The last command starts Tholos at http://127.0.0.1:7070. Open it, then:

  1. In Settings, press Detect and add the model it finds.
  2. Load the Research desk team from Starter teams.
  3. Choose that model for its three assistants: Scout, Analyst and Writer.

Scout checks Hacker News, arXiv and Hugging Face Papers every two hours, Analyst scores what it finds and Writer sends you a brief on Friday afternoon.

Frequently asked questions

What does Tholos do?

Tholos is an Apache-2.0 workspace where AI assistants carry out scheduled tasks, share tables and notes on your computer, and ask for approval under your rules. Its task model, Tholos-2B, is a 1.6 GB file that runs on an ordinary processor. You can also connect another local or hosted model through an OpenAI-compatible API.

Can AI assistants run on my own computer without the cloud?

Yes. Every table, note and schedule stays on your machine, and the task model Tholos-2B runs on a standard processor (CPU). On a CPU it completes 134 of 160 Tholos-Bench workspace tasks, so a whole assistant team can run on a laptop with no graphics card and no cloud account.

What is Tholos-2B?

Tholos-2B is a model with about two billion parameters, trained further from MiniCPM5-2B on generated task runs whose results passed the scenario checks. On one Kaggle T4 graphics card it completes 137 of 160 Tholos-Bench tasks, 25 more than the model it was trained from. You can download it from Hugging Face under Apache-2.0.

How much does Tholos cost?

Nothing. Tholos, its model Tholos-2B, its training runs and its benchmark are all free and open source under Apache-2.0. Two commands install the app and pull the 1.56 GB model file. There is no account, no subscription and no paid plan. With the default local Tholos-2B setup, the model and your workspace data stay on your computer.

How do several AI assistants share one workspace?

Each assistant has a role, a model, a schedule and permitted tools. They read and write the same tables and notes, and Tholos records every change and stops an assistant from overwriting a row before reading its latest version. The Research desk starter team shows the pattern with three assistants: Scout, Analyst and Writer.

How does Tholos keep assistants under control?

Rules allow an action, require your approval or deny it. Opening a web page requires approval by default, and the proposed address or table change is shown for review in the browser or on Telegram. On 12 Tholos-Bench tasks with misleading instructions hidden in the source material, Tholos-2B did the right thing 10 times.

What is Tholos-Bench?

Tholos-Bench is an open benchmark with 160 workspace tasks in 14 categories, scored on the final state of the workspace after the assistant finishes. Anyone can download it from Hugging Face and run it against their own model server. Tholos-2B passes 137 tasks; its starting point, MiniCPM5-2B, passes 112.