Four threads,
one open agenda.
Pure-PyTorch ports of frontier architectures. Reasoning and agentic-coding distillation from closed models onto small open weights. Datasets we built because nobody published them at this size. And the tools — claire, taskgen, claude-code-proxy — that came out of our own loop.
Architecture portability
Frontier weights locked behind custom CUDA kernels are not really open. We re-implement architectures (Nemotron-H, Cascade-2) in pure PyTorch so the original weights load with bitsandbytes, train with QLoRA, and run on a single consumer GPU.
Distillation at home
Reasoning and agentic-coding capability distilled from closed frontier models onto small open weights — RTX 5090, QLoRA, the whole pipeline. Dataset, training script, eval and weights all published together.
Datasets, public and raw
Six tasklist-* corpora generated by taskgen across math, code, science, creative writing and conversation. Apache-licensed code, public datasets, no quality gates. "Unfiltered" is in the name on purpose.
Tools we needed
claire is the flagship — a Rust-native, local-first coding agent that runs against any provider you have keys for. claude-code-proxy, taskgen and nanoAttnRes round out the loop. Built because we needed them; open because someone else might too.
First in line for claire.
One letter every other Tuesday — and a single dispatch on the day claire ships with the install line. What we shipped, what we read, the one thing we got wrong. No hype, no roadmap teasers. Cancel from any line.