Kisoku, a language model we trained from scratch.
A free, open 1.6B model that one person trained from scratch. It ties Meta's Llama 3.2 1B on ten benchmarks with about 18 times less training data, and it runs offline on your laptop.
Who are you? I'm Kisoku, a small open language model built from scratch by 0ARCH. I run on a laptop and handle everyday stuff like answering questions, writing, and code.
Explain why the sky is blue in two sentences. The sky is blue because sunlight is scattered in all directions by the air molecules, and shorter wavelengths of light (blue) are scattered much more than longer ones (red). This is called Rayleigh scattering.
Write a Python function that checks whether a number is prime. ```python def is_prime(n): if n < 2: return False for i in range(2, int(n**0.5) + 1): if n % i == 0: return False return True ``` This function checks divisibility up to the square root of `n`, which is an optimization over checking all the way up to `n`. If `n` is divisible by any number in this range, it's not prime. Otherwise, it's prime. You can use this function like this: ```python print(is_prime(25)) # False print(is_prime(23)) # True ```
- parameters
- 1.6B
- training tokens
- 524B
- less data than Llama 3.2 1B
- 18x
- recommended download
- 1.7 GB
- words a second on an M5 Max
- ~150
How it compares
We ran Kisoku and Meta's Llama 3.2 1B through the same test harness on the same machine. Out of ten benchmarks, each wins five. Kisoku takes the reasoning and math tests, Llama takes the knowledge and commonsense ones. The difference is the training data: Llama 3.2 1B is reported at about 9 trillion tokens, Kisoku saw about 524 billion, roughly 18 times less. So this is a result about getting more out of less data, not a claim to the best small model. Qwen2.5 1.5B and SmolLM2 1.7B are ahead of both on most of these tests.
5
Kisoku 1.6B
wins
5
Llama 3.2 1B
- Kisoku 1.6B: 15.0
GSM8K
Llama 3.2 1B: 5.8 - Kisoku 1.6B: 34.2
MMLU
Llama 3.2 1B: 31.3 - Kisoku 1.6B: 64.4
ARC-Easy
Llama 3.2 1B: 61.8 - Kisoku 1.6B: 40.1
ARC-Challenge
Llama 3.2 1B: 36.9 - Kisoku 1.6B: 28.6
BBH
Llama 3.2 1B: 28.3 - Kisoku 1.6B: 73.0
PIQA
Llama 3.2 1B: 74.9 - Kisoku 1.6B: 56.8
WinoGrande
Llama 3.2 1B: 60.5 - Kisoku 1.6B: 57.8
HellaSwag
Llama 3.2 1B: 64.2 - Kisoku 1.6B: 14.6
HumanEval
Llama 3.2 1B: 18.9 - Kisoku 1.6B: 23.1
TriviaQA
Llama 3.2 1B: 40.7
Long context: on RULER (13 tasks, averaged) Kisoku scores 76.1 at 4K tokens, 58.8 at 32K, 56.7 at 64K and 43.2 at 128K. Llama 3.2 1B scores 73.5, 56.7, 49.2 and 43.1. IBM Granite 4.0 1B and Qwen3.5 are ahead of Kisoku at every length.
Run it in three steps
Kisoku runs in Ollama, a free app that runs AI models on your own computer. The commands below are for the Terminal app on a Mac, or a Linux terminal.
- 1
Install Ollama
Download it from ollama.com and open it once.
- 2
Download Kisoku
This grabs the recommended 1.7 GB chat model and its setup file into a new folder.
$ mkdir kisoku && cd kisoku$ curl -LO https://huggingface.co/0arch-io/kisoku-1.6b-gguf/resolve/main/kisoku-1.6b-chat-sft009-Q8_0.gguf$ curl -LO https://huggingface.co/0arch-io/kisoku-1.6b-gguf/resolve/main/Modelfile
- 3
Start chatting
The first command sets it up, the second opens a chat. Type /bye to leave.
$ ollama create kisoku -f Modelfile$ ollama run kisoku
What it does well
- Everyday questions and conversation.
- Short writing: emails, summaries, rewrites, poems.
- Simple code, like a small Python function with an example.
Where the chat model is rough
- The chat model is a preview. In long conversations it can lose track of what it said earlier.
- It states wrong answers as confidently as right ones, so check anything that matters.
- A typo in a long request can throw it off.
- It sometimes refuses things it could just do.
- It has no internet access and knows nothing about current events.
Everything is open
The weights, the code, the technical report and every eval answer are public under Apache 2.0.
The base model is the finished release. The chat model is still a preview: it works, but expect the rough edges listed above.
Base weights
The pretrained model, with branches for stage 1, stage 3 and the two long-context checkpoints.
Chat model (preview)
Pass 9 of supervised fine-tuning. This is the one the steps above download.
GGUF for llama.cpp and Ollama
Q8_0 (1.7 GB, recommended), Q4_K_M (1.0 GB) and F16 (3.2 GB), for both the chat and base models, plus the Modelfile.
Code and technical report
The training code and the full write-up of how Kisoku was built.
Eval results
Every benchmark answer, sample by sample, so you can check the scores yourself.
Apache 2.0 license
Use it, change it and build on it, commercially too.
Spec sheet
- Size
- 1.6 billion parameters (1,600,749,056)
- Trained on
- About 524 billion tokens of text, including long-context phases (a token is roughly three quarters of a word). Trained on a TPU v4-32 from Google's TPU Research Cloud, a free grant.
- Context
- Trained to 64K tokens, about 128K with YaRN. The Ollama setup below uses 8,192 tokens by default.
- Download
- 1.0 to 3.2 GB
- Speed
- About 150 words a second on a MacBook Pro with an M5 Max chip
- Language
- English
- License
- Apache 2.0 for the weights, code and report
Questions
What is the best small open source language model that runs offline on a laptop?
It depends on the job, and Kisoku is one honest option. It is a 1.6 billion parameter model we trained from scratch, free under the Apache 2.0 license, and it runs offline on a laptop through Ollama with a 1.0 to 3.2 GB download. It is good at everyday questions, short writing and simple code. It ties Llama 3.2 1B on our ten-benchmark suite, but Qwen2.5 1.5B and SmolLM2 1.7B score higher on most tests, so if you only want the strongest model at this size, try those too.
How does Kisoku compare to Llama 3.2 1B?
On ten benchmarks run through the same harness on the same machine, each wins five. Kisoku is ahead on GSM8K, MMLU, ARC-Easy, ARC-Challenge and BBH (reasoning and math). Llama 3.2 1B is ahead on PIQA, WinoGrande, HellaSwag, HumanEval and TriviaQA (knowledge and commonsense). Kisoku got there on about 524 billion training tokens, against the roughly 9 trillion reported for Llama 3.2 1B, about 18 times less. As a chat model Llama 3.2 1B Instruct is still ahead: on a 297-conversation test a strict judge accepted 56.7% of its turns, against 48.9% for the Kisoku chat preview.
Does Kisoku work without internet?
Yes. Once the model file is downloaded, Kisoku runs entirely on your computer in Ollama. Nothing you type is sent anywhere, and it has no internet access of its own.
How big is Kisoku?
1.6 billion parameters. The download is 1.0 to 3.2 GB depending on the version, and the recommended one is about 1.7 GB. It writes about 150 words a second on a MacBook Pro with an M5 Max chip.
Is Kisoku free to use?
Yes. The weights, the code and the technical report are all under the Apache 2.0 license, which allows commercial use. The details are on the Hugging Face model page.