OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

TL;DR

OpenResearcher is a fully open recipe for training long-horizon deep-research agents without depending on commercial search APIs: a local ~11B-token text corpus + a self-built dense/BM25 retriever stands in for Serper/Google, and GPT-OSS-120B (with native browser tools) is used as the trajectory synthesizer. The pipeline produces 96K research trajectories of 100+ turns each, which SFT-tunes Nemotron-3-Nano-30B-A3B-Base into a 30B-A3B research agent that hits 54.8% on BrowseComp-Plus — claimed above GPT-4.1, Claude-Opus-4, Gemini-2.5-Pro, DeepSeek-R1, and Tongyi-DeepResearch. The release covers data, weights, training code (Megatron-LM fork), the retriever+corpus, an eval framework, and all eval logs. The thesis is that the bottleneck for open research agents is not the agent loop but the cost and non-determinism of the search environment, and that an offline retriever + corpus is enough environment to train a frontier agent.

Key claims

A local 11B-token corpus + self-built retriever (BM25 or Qwen3-Embedding-8B dense) replaces commercial search APIs as the agent’s information environment, eliminating per-query API spend ($0.001–0.01/query at Serper-tier pricing), rate limits, and non-determinism [§Introduction, README §Features].
The pipeline synthesizes 96K long-horizon deep-research trajectories (100+ turns each) using GPT-OSS-120B as the driver LLM with native browser tools against the offline retriever [README §Features].
SFT on these trajectories lifts Nemotron-3-Nano-30B-A3B-Base from 20.8% → 54.8% on BrowseComp-Plus, surpassing the reported figures for GPT-4.1, Claude-Opus-4, Gemini-2.5-Pro, DeepSeek-R1, and Tongyi-DeepResearch on the same benchmark [README §Introduction].
Within the open-source landscape, OpenResearcher is positioned as the first release that ships all four of: weights, full SFT trajectory corpus, training code, and a reproducible training environment (retriever + corpus); Search-R1 lacks the trajectories, Tongyi DeepResearch lacks code+env, MiroThinker lacks the environment [§Comparison table].
Strong results are claimed across a broader benchmark suite — BrowseComp, BrowseComp-Plus, GAIA, xbench-DeepSearch — using a unified evaluation harness shipped with the release [README §Deep Research Benchmark Results, §Benchmark OpenResearcher].
The pipeline has been adopted by NVIDIA’s Nemotron family of models and powers deep-research trajectory generation in NVIDIA’s NeMo Data Designer [README §News, 2026.2.12 entry].

Method

The pipeline is structured around an offline-first agent environment. A dedicated ~11B-token text corpus (released on HuggingFace as OpenResearcher-Corpus) plus a deployable retriever service (BM25 for compute-cheap baseline, Qwen3-Embedding-8B dense for accuracy) replaces a live web search API. The agent loop is a standard ReAct-style trajectory of thought / search / browse / observation steps; for trajectory synthesis, GPT-OSS-120B is the driver, configured with native browser tools (vLLM recipe) that route through the local retriever rather than the public web. Trajectories are filtered for length, correctness signals, and tool-call validity, yielding the 96K OpenResearcher-Dataset.

Training is plain SFT (no RL stage) on Nemotron-3-Nano-30B-A3B-Base, run via a Megatron-LM fork released under the openresearcher branch. Inference uses vLLM tensor-parallel servers (TP=2 across 4 GPUs is the documented default) coupled to either the local dense retriever (used for BrowseComp-Plus, which provides its own corpus and embeddings) or the Serper API (used for BrowseComp, GAIA, xbench — the offline corpus is specific to BrowseComp-Plus and does not cover those benchmarks’ web environments). The eval framework ships unified runners for all four benchmarks with an OpenAI-API-based scorer.

Results

Headline number: 54.8% on BrowseComp-Plus by the 30B-A3B SFT model trained on 96K synthetic trajectories, up from 20.8% base [README §Introduction]. This places it above the closed-source models GPT-4.1, Claude-Opus-4, Gemini-2.5-Pro and the open-source DeepSeek-R1 / Tongyi-DeepResearch on the same benchmark per the project’s reporting. Trajectory generation cost is reported as substantially lower than API-based synthesis: search/scrape API spend is zero (offline retriever), only the GPT-OSS-120B inference cost remains. The HuggingFace dataset crossed 11K downloads and ranked top-3 on HF Trending Datasets in Feb 2026 [README §News, 2026.2.25]. The arXiv paper hit #2 on HF Daily Papers on 2026-03-25 [README §News].

Why it’s interesting

Three things to flag for Luma. First, this is the cleanest existing open datapoint that the bottleneck for training research agents is the environment, not the model: replacing Serper with a local retriever-over-corpus drops the marginal cost of a 100-turn trajectory toward zero and removes the run-to-run nondeterminism that makes large-scale RL/SFT data pipelines flaky. The same logic applies to any generative agent whose tool calls hit external APIs — video-editing assistants, asset-search agents, multimodal research helpers. Second, the contrast with MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling is instructive: MiroThinker reaches 47.1% on BrowseComp at 72B with a 3-stage SFT→DPO→GRPO pipeline over a live (API-based) environment; OpenResearcher reports 54.8% on the (different but related) BrowseComp-Plus at 30B-A3B with SFT-only over an offline environment. The two papers do not run on the same benchmark — BrowseComp-Plus is the retriever-isolating variant of BrowseComp from arXiv:2508.06600 — so headline numbers don’t compare directly, but the recipe is meaningfully simpler. Third, this is another instance of Synthetic Training Data‘s pattern: a strong frontier synthesizer (GPT-OSS-120B) + a verification surface (does the trajectory end in a correct answer against the corpus?) generates training data that beats live-API recipes, with the curation gate doing load-bearing work.