Bronze
Back the work
$10 / month
- Your name on our supporters list
- A vote on what we build next
- A thank you in the build log
Open-weight model
Kepler
An open AI model for space math. Ask it an orbit question and it shows short work that ends in one number you can check. It runs fully offline on a small desktop, for free.
- Field
- Space math
- Runs
- Fully offline
- Built on
- Qwen3 8B
- License
- Apache-2.0, free
Kepler
Kepler is an open AI model for space math. That field is called orbital mechanics, the math of how things move in space, like a satellite circling Earth. Ask Kepler one of those questions and it shows a short chain of work, then ends with one clear number you can check.
What it can do
It answers word problems about orbits, speeds, and travel between planets. It is built on Qwen3 8B, a strong small model, and taught with 600 worked space problems. A checking program graded every one of those problems, so the model learned from answers that were known to be right.
The teaching also fixed a bad habit. The base model would think forever and run out of room before giving an answer. Kepler answers every time, never gets cut off, and uses about a third of the words big cloud models spend on the same question.
How well it works
We score it on a held-out test of 44 space problems it never saw in training. A program checks each answer to within 2 percent. The best build gets 88.6 percent right, and the full ladder is in the table above.
We also ran the same test against small frontier cloud models, with the same rules for everyone. Kepler got 84.1 percent on that run. Claude Haiku 4.5 got 97.7 and Gemini Flash-Lite got 95.5. That is the honest read: the cloud models are still better at the hardest multi-step problems. But Kepler runs on your own desk, free, private, and offline, and it answers in 166 words of work on average where the cloud models spend close to 500.
Where it falls short
Two kinds of problems trip it up: two-burn transfers, the maneuvers that move a craft between two orbits, and working backward from an orbit time to a height. Treat its answers there as a first draft and check the number. We say this plainly because the test set that found these gaps is public too.
How to run it
Download the Q8_0 build, an 8.2 GB file, and run it with llama.cpp on a Spark-class desktop, a small AI machine with 128 GB of memory. LM Studio and Ollama load the same file with no extra setup. If memory is tight, the smaller builds still work, they just miss more of the hardest problems.
Install
huggingface-cli download Orionfold/Kepler-GGUF model-Q8_0.gguf --local-dir ./models/keplerllama-cli -hf Orionfold/Kepler-GGUF:Q8_0Use it
# A small server on your own machine. Anything that talks to
# OpenAI-style APIs can now talk to Kepler instead.
llama-server -m ./models/kepler/model-Q8_0.gguf \
-c 4096 -ngl 99 --host 0.0.0.0 --port 8080
from llama_cpp import Llama
llm = Llama(
model_path="./models/kepler/model-Q8_0.gguf",
n_ctx=4096, n_gpu_layers=99, chat_format="chatml",
)
out = llm.create_chat_completion(messages=[{
"role": "user",
"content": "A satellite circles Earth 550 km up. What is its "
"orbital period in minutes? Give your final answer "
"as \\boxed{value unit}.",
}])
print(out["choices"][0]["message"]["content"])
Specs
- Base model
- Qwen/Qwen3-8B
- Taught with
- 600 worked space problems, every one checked by a program
- Format
- GGUF (ready to run)
- Builds
- Q4_K_M · Q5_K_M · Q6_K · Q8_0
- Best build
- Q8_0 (8.2 GB, keeps almost all the quality)
- License
- Apache-2.0 (free to use)
Benchmarks
| Build | Size | Right answers (44-question test) |
|---|---|---|
| Q4_K_M | 4.7 GB | 75.0% |
| Q5_K_M | 5.5 GB | 75.0% |
| Q6_K | 6.3 GB | 84.1% |
| Q8_0 | 8.2 GB | 88.6% |
Used in the open
Live counts from HuggingFace, refreshed when the site builds. Built and maintained in the open by Orionfold.
- 109
- Downloads · last 30 days
Sponsor Kepler
Back this work with a monthly tier. Your support moves your requests up the list, and Gold or Platinum earns a badge on the roadmap item you back.
Silver
Get a say
$25 / month
- Everything in Bronze, plus:
- Your requests join the priority line
- Early access to new versions
Gold
Move it up the list
$50 / month
- Everything in Silver, plus:
- High priority on your feature and fix requests
- A Gold badge on the roadmap item you back
- A monthly call to plan the work
Platinum
Shape the roadmap
$100 / month
- Everything in Gold, plus:
- Top of the line, we work your requests first
- A Platinum badge on the roadmap item you back
- A direct line to us and a hand in the roadmap
Need something specific? Send an enquiry from the roadmap.
Keep exploring
Book
AI Research on NVIDIA DGX Spark
Real notes from doing AI research on one desktop. The NVIDIA DGX Spark is a small machine with huge power (petascale means it runs about a quadrillion math steps a second), so you can push local AI further with no cloud needed. Every lesson is backed by code that runs.
Story
The story of the run that built it
I built Kepler, a space-math model, on one desktop. Three cheap checks told me the big training run would change nothing. I ran it anyway, and proving them right was the win.