Hacker News surfaced nanocode as an unusually explicit recipe for training a Claude Code-style model without a hyperscaler-sized budget. The project describes itself as an end-to-end library for tokenizer training, pretraining, synthetic data generation, agentic supervised fine-tuning, and DPO with Constitutional AI, all implemented in pure JAX and aimed primarily at TPU workflows.

The part that stood out to HN readers is the economics. According to the README, a 1.3B-parameter nanocode-d24 run can be reproduced in about 9.3 hours on a TPU v6e-8 for roughly $200. Smaller variants are cheaper still: the 477M-parameter d20 target is listed at about $34, and the 135M-parameter d12 configuration at about $3. Those are not consumer laptop numbers, but they are low enough to shift the conversation from “only big labs can try this” to “small teams can run serious experiments if they have the right infra.”

The repository is also notable for not stopping at base model training. It includes scripts and datasets for coding rollouts, tool use, long-context scenarios, preference optimization, and a CLI agent. The synthetic data pipeline pulls in multiple coding instruction sources, then adds critique and restyle passes so the project can move from raw datasets to agent-like behavior instead of just next-token prediction.

There are obvious caveats. The author is transparent that nanocode is optimized for TPUs, and the workflow expects Google Cloud administration, SSH into TPU VMs, and JAX-specific tuning. This is still a research-grade path, not a one-click product. But that is also what makes it interesting: it exposes the actual moving pieces behind a coding model pipeline instead of hiding them behind an API.

The HN discussion is at news.ycombinator.com/item?id=47649742, while the implementation is documented in the nanocode repository. For anyone trying to understand what open agent training looks like in practice, this is one of the more concrete blueprints currently circulating in the community.

#jax

Hacker News Highlights Nanocode, a JAX/TPU Path to Train a Claude Code-Style Model for About $200