NVIDIA Is Giving Away the Agent Trajectories Everyone Else Is Paying For

The scarcest input in AI right now is not compute, and it is not text scraped off the open web. It is the recorded log of an agent doing something hard and multi-step: opening a repository, running a failing test, reading the traceback, editing three files, running the test again, and either fixing the bug or failing in an instructive way. Those logs are called trajectories, they are what separate a chatbot from a software engineering agent, and nearly every lab that has them treats them as a trade secret. NVIDIA has spent the past five months uploading millions of them to Hugging Face under a license that lets anyone — including its customers, and including its competitors — train on them commercially, for free.

The scale is not subtle. nvidia/Open-SWE-Traces, refreshed again this week, holds 207,489 agent trajectories drawn from roughly 20,000 real-world pull requests across nine programming languages: Python, Go, TypeScript, JavaScript, Rust, Java, PHP, C and C++. They were collected inside three agent harnesses — OpenHands, SWE-agent and mini-swe-agent — and run to 46.6 GB across 567,824 rows. Sitting next to it are SWE-Zero-openhands-trajectories (318,115 trajectories over 118,092 issues), the execution-verified SWE-Hero set (roughly 34,000 trajectories from 11,766 curated issues), and Nemotron-SFT-Agentic-v2, a general tool-use corpus of about 1.2 million single-turn, multi-turn and multi-step trajectories, including 278,880 customer-service-style conversations across 838 domains and 6,977 web-search trajectories that mostly require 10 to 30 search calls apiece.

Every one of those datasets ships under Creative Commons Attribution 4.0 — not the NVIDIA Open Model License, not a research-only carve-out. The dataset cards state that the data is "ready for commercial/non-commercial use," and the source repositories were filtered down to MIT, Apache-2.0, BSD-2-Clause and BSD-3-Clause licenses before any agent was pointed at them. Open-SWE-Traces alone was downloaded 10,111 times last month.

The researchers are explicit about why this matters. "The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data," write Wasi Uddin Ahmad, Nikolai Ludwig, Somshubra Majumdar and Boris Ginsburg in the accompanying paper, published June 14 as arXiv 2606.16038. The team validated the corpus by fine-tuning the Qwen3-30B-A3B series on it; the best resulting model resolved 61.7% of SWE-bench Verified, 57.1% of SWE-bench Multilingual and 36.8% of SWE-bench Pro — respectable numbers for a 30B-class open model trained entirely on data anyone can download.

An American Chip Company, Chinese Teachers

Here is the wrinkle nobody at NVIDIA has drawn a box around. The teacher models that generated this data are almost entirely Chinese open-weight releases. Open-SWE-Traces uses a dual-mode synthesis recipe: MiniMax-M2.5 produces trajectories with explicit thinking traces, while Qwen3.5-122B-A10B supplies the non-thinking traces. An update logged on the card this week added trajectories generated by DeepSeek-V4-Flash and Qwen3.6-27B. SWE-Zero and SWE-Hero were both synthesized with Qwen3-Coder-480B-A35B-Instruct. The issue statements themselves come from community sets like SWE-rebench-V2, SWE-Gym, R2E-Gym and SWE-Fixer-Train-110K.

So the pipeline runs like this: Alibaba, MiniMax and DeepSeek publish open weights; NVIDIA burns its own GPUs distilling those weights into hundreds of thousands of agent trajectories; the trajectories go out under CC BY 4.0; and whoever downloads them trains their next agent on — NVIDIA hardware. No American frontier lab offers an equivalent permissively-licensed teacher, because their terms of service forbid precisely this kind of distillation.

NVIDIA has been open about the broader posture. In a March post on Hugging Face, the company said it has "shared more than 2 petabytes of AI-ready training data across more than 180 datasets and 650+ open models. And we are just getting started." A July follow-up, Data for Agents, made the agentic case bluntly: "An agent that cannot recover from a broken API call, or a workflow it has never seen, is not really an agent. It is an autocompleter with tools." It closes on a line that reads like a thesis statement: "The scarce resource in AI is not tokens. It is trust between organizations."

Why It Matters

Agentic post-training data is the current bottleneck input, and everyone in the industry knows it. Data-labeling and RLHF vendors have built large businesses on the premise that high-quality agent trajectories are expensive, artisanal, and worth paying a premium for. Frontier labs guard their internal trajectory stores the way they once guarded model architectures.

NVIDIA has no reason to defend that scarcity. It does not sell trajectories. It sells the machines that consume them. Every downstream team that can skip a six-month, seven-figure data-collection program and go straight to a fine-tuning run is a team that rents GPU hours sooner and at greater scale. Commoditizing your complement is the oldest move in the platform playbook, and here the complement is not just model weights but the data layer underneath them.

The company even signals that it understands the asymmetry. The Data for Agents post cites NVIDIA VP of applied deep learning research Bryan Catanzaro on the idea that "every company is built around a secret" — a workflow, corpus, or customer pattern competitors do not have. Trajectory data is emphatically not NVIDIA's secret. Which is exactly why it can afford to give it away, and to squeeze the margin out of everyone whose secret it is.

What to Watch

Two things. First, whether the market for generic agent trajectories reprices. If a free 200k-trajectory corpus takes a 30B open model to 61.7% on SWE-bench Verified, the floor for what a buyer will pay for undifferentiated software-engineering traces has just moved, and the vendors will need to retreat further up the specialization curve. Second, whether the Chinese-teacher dependency starts drawing attention. NVIDIA's fastest route to permissively licensed agentic data currently runs through Hangzhou and Shanghai, and that is a supply chain — legal, reputational, and potentially regulatory — that nobody has stress-tested in public yet. Watch whether the next release swaps in a Nemotron teacher, and what that does to the resolve rates.