hyperresearch vs OpenResearcher
Deep-research harness for Claude Code: a 16-step, tier-adaptive pipeline with adversarial critics, cite-checking and 250+ sources per run, every source kept in a persistent markdown+SQLite vault. — versus — TIGER-AI-Lab's fully open deep-research recipe: 96K long-horizon trajectories (adopted by NVIDIA Nemotron), a 30B-A3B model hitting 54.8% BrowseComp-Plus, training code and eval harness.
openresearcher is a training recipe for a deep-research model; hyperresearch gets there with harness engineering on a frontier model - no training, more steps.
| hyperresearch | OpenResearcher | |
|---|---|---|
| Stars | 3.4k | 1.2k |
| Forks | 336 | 122 |
| Language | Python | Python |
| License | MIT | — |
| Last activity | 7 days ago | 3 months ago |
| Topics | agents, rag, skills | training |
| Curated connections | 5 | 4 |
hyperresearch — the curator's take
The most serious open deep-research pipeline you can run inside Claude Code: width sweep, contradiction graph, parallel depth investigators, four adversarial critics and a tool-locked patcher that can only apply surgical edits. Scholarly search fans out to OpenAlex/Crossref/CORE/EDGAR/FRED and it fetches legal OA full-text for paywalled papers instead of citing abstracts. Use it for reports you would otherwise pay a research analyst for; the vault means the second run on a topic starts warm. Not for quick lookups - even the 5-step fast path is minutes and the full pipeline is 30 min to hours of Claude Code time, plus your Max quota. Locked to Claude Code (uses the Skill tool for step loading); last30days-skill is the fast social-signal alternative, openresearcher is the trained-model route.
OpenResearcher — the curator's take
The open counterpoint to closed Deep Research products — and the data is the crown jewel: 100+-turn research trajectories distilled from GPT-OSS-120B over a self-built 11B-token retriever corpus (no search-API bills at generation scale), good enough that NVIDIA folded it into Nemotron 3 Ultra. The 30B-A3B model beats GPT-4.1, Claude-Opus-4 and Gemini-2.5-Pro on BrowseComp-Plus. Reproducing anything is a real commitment: the setup assumes 8×A100, training lives in a Megatron-LM fork, and the local retriever needs Java + tevatron. ⚠ No LICENSE file in the repo — clarify terms before commercial use. Pick DeepDive for the KG-synthesis + multi-turn-RL recipe; OpenResearcher for large-scale SFT distillation with everything — data, model, eval — actually released.