This week's free AI tools: six you can run yourself
The six free tools from this week's Run It Yourself episodes, Mon Nov 2 to Sat Nov 7. Test-box numbers are from our 8-core CPU-only box, not typical performance.
Groot-19B
Groot-19B-A3B-261023 from GrootTeam is free, open weights under Apache 2.0, with about 19B total and about 3B active parameters per token; the official Q4_K_M GGUF is an 11.7 GB download, and its makers say it runs on a 16 GB card like an RTX 4060 Ti. On our box with no GPU, just 8 CPU cores, it answered our question correctly at 7.9 tokens/s. Try it (needs llama.cpp): llama-server -hf GrootTeam/Groot-19B-A3B-261023-GGUF:Q4_K_M, then open localhost:8080.
Sources
Local Code
A free, open-source (GPL-3.0) AI agent app for the Mac from the team behind Draw Things; on-device models cost $0 per token, and the smallest, Qwen 3.8 27B, needs a Mac with 16 GB of memory. If you pick a cloud model, your prompts and files go to that provider. Try it: download Local Code from getlocalcode.app, then pick a model that fits your Mac's memory.
Sources
- Local Code official site (getlocalcode.app)getlocalcode.appPrimary
- Draw Things: Local Code release postreleases.drawthings.aiPrimary
- drawthingsai/local-code-community (GitHub)github.comPrimary
- MacRumors: Local Code launches as free, private AI assistant tool for Macmacrumors.com
- iThinkDiff: Local Code is a free, open-source AI agent for Macithinkdiff.com
Bonsai 2 2B
From PrismML, free under Apache 2.0, with every weight -1, 0 or +1; bonsai-2:2b is a 911 MB download in Ollama. On our CPU-only test box it ran at about 36 tokens/s using about 1.1 GB of RAM. Try it: install Ollama (ollama.com), then run: ollama run bonsai-2:2b
Sources
Kisoku 1.6B
From 0ARCH, trained from scratch by one person, Joseph Rodriguez, and free under Apache 2.0; the recommended chat download is 1.7 GB and runs in Ollama. It is a preview chat model that, per its maker, can state wrong answers confidently; on our CPU-only test box it ran at about 40 tokens/s using about 2.5 GB of RAM. Try it: in a new folder, download kisoku-1.6b-chat-sft009-Q8_0.gguf and Modelfile from huggingface.co/0arch-io/kisoku-1.6b-gguf, then run: ollama create kisoku -f Modelfile, then: ollama run kisoku
Sources
MiniCPM5-1B
From OpenBMB, about 1B parameters, free under Apache 2.0, built for on-device use, and a 688 MB download on Ollama. On our CPU-only test box it got the math right but missed a Spanish translation request in our recording (asked fresh, outside the recording, it translated correctly), at about 45 tokens/s and about 0.8 GB of RAM. Try it: ollama run openbmb/minicpm5
Sources
LFM3-VL-2B
A vision-language model from Liquid AI with about 2 billion parameters that reads text and images (documents, charts, app screens); open weights, free to use under the LFM Open License unless your company makes $10M+ a year, 1.5 to 1.7 GB on Ollama. On our CPU-only test box it read a receipt we made correctly; one receipt is one test, so check its answers. Try it: ollama run lfm3-vl, then type your question followed by the image path, e.g. What is the total? ./receipt.png
Sources