Quark
Our first dataset, built for concise chain-of-thought reasoning and token efficiency. It packs more reasoning steps into the same output length, so models think further per token instead of spending tokens to think.
Hugging FaceWhat we’ve built so far.
Three kinds of work make up the lab: fine-tuned models, the tools that train them, and the datasets that shape them.
Models.
Three models make up a scaling system, named for what each class orbits: a satellite, a planet and a star. Kuiper sits apart as a specialist.
- Hugging Face
Deimos R1
Sharper reasoning. Fewer tokens.
Our smallest model and our reasoning specialist. It carries the Deimos line’s short, token-efficient chain of thought into a full release.
- Hugging Face
Europa B1
More intelligence per parameter.
Our medium-sized model. We use it to see how far intelligence scales per parameter, and to push the reasoning techniques from Deimos further.
- Hugging Face
Ganymede A1
More correct answers. Less reasoning overhead.
Our flagship. It takes what we’ve refined on the smaller models and tests it on real tasks, where efficient reasoning matters most.
- Hugging Face
Kuiper R1
A trace-inversion model. Give it another model’s output and it reconstructs the reasoning behind it, which makes results easier to audit and interpret.
Tools.
The apparatus that generates, compresses and runs the training behind the models.
Hadron
A distillation framework built around NousResearch’s AutoReason tournament refinement. One teacher answers, critiques, revises, synthesizes and ranks its own work in a blind vote, and the loop repeats until “do nothing” wins twice. The labels that come out beat the teacher’s own single-shot answers, and every role leaves a full reasoning trace you can use for process-supervised fine-tuning.
GitHubTokamak
Pulls reasoning traces out of LLM conversations and compresses them into a compact stream of internal chain of thought and short outputs. It’s meant for making tight, high-signal training data from long, branching dialogues without losing the reasoning that got you there.
GitHubStellarator
A control plane for fine-tuning and reinforcement-learning runs on Tinker. Sandbox runs pass a structured pre-flight check before they scale up, with cost projections, budgets and live alerts at every step. A Rust supervisor handles per-job polling and websocket fan-out, and a built-in research subsystem cites Hugging Face papers, arXiv and code examples for each run.
GitHub
Datasets.
The data that shapes how the models reason.
Photon
A family of supervised fine-tuning datasets for hardening models around compliance. Instead of bolting guardrails on at inference time, Photon builds decision-tree-structured compliance reasoning into the weights during fine-tuning, so compliant behavior is a capability of the model and not an add-on. Each domain variant is called an isotope. Isotopes share a schema and a training approach but target different regulatory and security areas.
