NanoResearch is OpenRaiser’s MIT-licensed nine-stage engine that takes a research idea to a complete LaTeX paper by running the experiments itself: the pipeline submits generated code to local or SLURM GPU clusters, parses the training logs that come back, and writes every table and figure from those outputs.
It is the idea-to-paper column that binds the manuscript to actual runs rather than sandbox simulations, and its resumable pipeline plus per-stage model routing make it the practical choice for anyone whose experiments need a GPU cluster, with the usual discount that the evidence is self-published.
What it is #
A Python CLI (github.com/OpenRaiser/NanoResearch, MIT, Python 3.10+, v1.0 released March 2026) with a bilingual Chinese-and-English README. The nine stages run from literature search (OpenAlex and Semantic Scholar) through experiment blueprints, code generation, GPU job submission, log parsing, figure generation, and LaTeX writing, with checkpoint-and-resume from any failed stage and a different model configurable per stage. Execution targets are a local GPU or a SLURM cluster with auto-submission; a Claude Code mode and a Feishu bot add chat and messaging surfaces to the CLI, and an Evo self-evolving pipeline is listed as a feature section in both READMEs. The README’s core claim: every number, table, and chart in the output paper comes from the pipeline’s own experiment outputs, not LLM fabrication.
Status #
Active: 1,340 stars, 92 forks, created 2026-03-17, last pushed 2026-10-09, as of 2026-10-11. The release history is thin (a single assets release in March 2026) despite the README’s v1.0 milestone, and a search for the name returns no Hacker News threads as of 2026-10-11.
Strengths #
- The evidence model is the differentiator: code goes to a GPU cluster, logs come back, and the paper binds to those logs, attacking the fabricated-results failure the citation checkers cannot see.
- Checkpoint-and-resume per stage plus per-stage model routing is engineering for the long, flaky runs actual experiments produce.
- Venue templates (NeurIPS, ICML, arXiv) and the claimed cost floor (as low as $0.5 of API credit per paper on DeepSeek) make it the cheapest member here per paper.
- Three surfaces (CLI, Claude Code mode, Feishu bot) suit both individual researchers and lab groups.
Cautions #
- The core claim is the README’s: this run found no independent evaluation, no benchmark integration, and no third-party account of a NanoResearch paper’s data integrity.
- A single March 2026 release asset against a v1.0 milestone suggests versioning lives in the README rather than in tagged releases.
- SLURM auto-submission runs generated code on shared infrastructure, a surface the README does not discuss sandboxing.
- The project publishes no paper of its own, so its design claims rest on documentation alone.
Pricing #
Free and open source under MIT; there is no paid tier, so pricing does not apply. Costs are your own GPU time (local or cluster) and LLM API calls; the README claims as low as $0.5 per paper with DeepSeek.
Compared to #
- AutoResearchClaw: the 23-stage rival with citation verification and human-oversight modes; NanoResearch is the GPU-cluster executor with no published citation-check layer.
- Sakana AI Scientist: the ancestor; Sakana’s experiments ran in managed sandboxes, NanoResearch’s go to schedulers you administer.
- Agent Laboratory: the human-assist ancestor with published quality scores; NanoResearch is fully autonomous with published costs instead.
Bottom line #
Recommended for researchers with GPU cluster access who want idea-to-paper runs where the paper’s numbers come from jobs the pipeline itself submitted. Not for anyone needing verified data integrity: the binding-to-logs claim is self-published, and the citation layer other pipelines ship is absent here.
Changes #
- 2026-10-11 - Created.
See also #
- AutoResearchClaw - the stage-count rival with the citation-verification layer NanoResearch lacks
- Sakana AI Scientist - the sub-family ancestor
- Agent Laboratory - the measurement-focused ancestor
- Automated Research Feature Matrix - the category comparison this note joins
References #
https://github.com/OpenRaiser/NanoResearch - repository, MIT license, 1,340 stars, last push 2026-10-09 (as of 2026-10-11)
https://raw.githubusercontent.com/OpenRaiser/NanoResearch/main/README.md - the Chinese main README: the actually-runs-experiments claim, nine stages, SLURM, Evo pipeline, Claude Code mode, Feishu bot (fetched 200, 2026-10-11)
https://api.github.com/repos/OpenRaiser/NanoResearch/releases?per_page=5 - the single assets release of 2026-03-19 (fetched 200, 2026-10-11)
https://hn.algolia.com/api/v1/search?query=NanoResearch - the search returning no threads (fetched 200, 2026-10-11)