Making KARL's innovations accessible to everyone.
KONASH is built by Kona Equity. We take breakthrough research from leading AI labs and ship it as open-source tools that anyone can use.
Why KONASH Exists
The best AI research shouldn't require a GPU cluster and a six-figure compute budget.
Databricks' KARL paper proved that RL-trained knowledge agents can match or exceed frontier models on grounded reasoning tasks. But KARL requires massive compute — hundreds of thousands of rollouts across multi-GPU clusters.
We believe the gains are algorithmic, not scale-dependent. KONASH packages the full KARL pipeline — agentic synthesis, OAPL training, parallel thinking — into a framework that runs with a single API key and costs under $5 per training run.
Our Principles
Open Source First
The full framework is Apache 2.0. Your model, your weights, your infrastructure. We never lock you in. The managed service adds convenience, not capability.
Research Fidelity
We implement papers faithfully. KONASH follows KARL's architecture section by section — OAPL loss, token masking, compression-as-skill, pass-rate filtering. No shortcuts.
Radical Accessibility
Two-minute setup. Interactive CLI. Pre-built indexes. If a PhD student on a free Colab tier can't run it, we haven't made it accessible enough.
Ship and Prove
Every claim comes with a benchmark, a reproduction script, and a trace viewer. We show results, not slides.
Credits
KONASH builds directly on the work of these teams and projects:
KARL
Databricks, 2026The architecture, training pipeline, and evaluation methodology that KONASH implements.
OAPL
Ritter et al., 2026Off-policy squared-advantage loss — the RL algorithm at KONASH's core.
Tevatron
HuggingFacePre-built BrowseComp-Plus embedding indexes for zero-setup evaluation.
Train your first agent today.
pip install konash. Two minutes to setup. Point it at your documents.