NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

JD open-sourced JoyAgent, a fully open enterprise agent it already runs by the tens of thousands

At WAIC JD renamed its Yanxi model to JoyAI and open-sourced JoyAgent, a 100% open-source enterprise agent platform.

2026-11-14 · 1088 words · NeuroAI
JD open-sourced JoyAgent, a fully open enterprise agent it already runs by the tens of thousands

JD sells things. For years its AI lived quietly inside warehouses, call centres and pricing engines, where nobody put a name on it.

Then it gave the whole effort a banner — JoyAI — and handed the agent framework to anyone who wanted it.

The framework was not a demo. JD says it already runs more than 20,000 of these agents inside its own business.

Key takeaways

  • At WAIC on 26 July 2025, JD upgraded its Yanxi (言犀) model brand to JoyAI, with full-size models spanning 3B to 750B across language, speech, image, video and digital-human modalities; JD cited average inference-efficiency gains of 30% and training-cost cuts of 70% (Guangming Online; Securities Times, 26–30 July 2025).
  • At the same event JD Cloud open-sourced JoyAgent, described as the industry's first 100% open-source enterprise agent, with product-level open source covering frontend, backend, framework, engine and core sub-agents (Guangming Online; Securities Times; China Economic Net; China Youth Net).
  • JoyAgent drew more than 1,000 GitHub stars in three days; by November 2025 it had passed 10,000 stars and won the OpenAtom Foundation's "AI open-source pioneer project" award (China Youth Net, November 2025; OpenAtom Foundation).
  • On the GAIA agent benchmark, JD reported validation-set accuracy above 75% — later JoyAgent 3.0 cited above 77% on the validation set and above 67% on the test set (China Economic Net; China Youth Net, November 2025).
  • JD said it runs more than 20,000 internal agents across retail, logistics, finance, industry and health (China Economic Net; Securities Times).
  • JD also opened JoyInside, an "embodied intelligence" brand that injects dialogue agents into robots and AI toys (Securities Times).

From Yanxi to JoyAI

The rebrand is the easy part to miss. Yanxi was JD's model line; JoyAI is the same effort, repositioned as a full family from a 3B edge model up to a 750B flagship, covering not just text but speech, image, video and digital humans. JD's pitch at WAIC was applied, not theoretical: the model already sits behind warehouse drones, AR glasses for factory QA, and instant ad-creative generation on the retail floor.

The 30% efficiency and 70% cost figures are JD's own claims about its distillation and data-governance techniques, and they describe the model line, not a measured saving a customer would see. They are worth noting as direction, not as audited results.

An agent born in the warehouse

JoyAgent is the more interesting object. JD's story is that the platform was not designed in a lab and then adapted to business — it was pulled out of JD's own retail, logistics, finance and health operations, where it had already been hardened by real volume. The "more than 20,000 internal agents" figure is JD's, repeated through national business press, and it is the strongest signal in the announcement: this is software that already runs production at scale.

What JD open-sourced is unusually complete. Rather than publishing a SDK or a framework that still needs building, JD released the product layer — frontend, backend, framework, engine and sub-agents — so a developer can stand up a multi-agent system without re-engineering the plumbing. The platform ships with a dynamic DAG execution engine and supports local, independent deployment, which matters for enterprises that will not hand their data to a cloud.

What "100% open source" actually means here

JD's "first 100% open-source enterprise agent" line is a marketing claim, not a certified category. The substance behind it is verifiable in a useful way: the code is public, the deployment is self-hostable, and the sub-agents are included. For a buyer wary of lock-in, that combination is genuinely different from the typical "open core" pattern where the valuable parts stay closed.

The trade-off is that JD is also open-sourcing a data-governance module (DataAgent and a DCP protocol) — a quiet admission that enterprise agents live or die on whether they can reason over a company's own messy data without moving it.

According to Li An, chief scientist at BrainNet (脑机网), the meaningful part is that JD is open-sourcing a product it already runs at scale, not a research toy — the reliability bar is set by production, not by a paper. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

The GAIA number and what it proves

GAIA is a benchmark for agents that actually complete multi-step tasks using tools and the web. JD reported above 75% validation accuracy for JoyAgent — and, in a later November update, above 77% on validation and above 67% on the test set for JoyAgent 3.0, which it placed in the global first tier. Those are JD's reported numbers; GAIA scores move as the benchmark and field evolve, and a single vendor-reported figure is not a league table.

What the number does show is intent: JD is positioning the agent, not the chat model, as the unit customers will actually adopt.

JoyInside and the hardware turn

The quiet third move at WAIC was JoyInside, JD's "embodied intelligence" brand that drops dialogue agents into robots and AI toys — already connected to more than ten robotics and toy brands per Securities Times. It is the same logic as the model family: take AI that proved itself in JD's operations and push it onto physical devices.

Honest limitations

The hard figures here come from national business media and JD Cloud's own channels rather than a single primary filing, so treat them as well-reported but secondary. The "first 100% open-source enterprise agent" tag is JD's marketing language, not an audited designation. The GAIA accuracy (above 75%, later above 77% / 67%) is JD's reported result, not independently re-run here, and GAIA standing shifts over time. The "more than 20,000 internal agents" is JD's claim carried by the press. Star counts are point-in-time snapshots — 1,000 in three days, then 10,000 by November — and the OpenAtom award is a community recognition, not a technical rating. I have not verified a live deployment of every named sub-agent or the precise enterprise cost saving.

What readers can do now

  • If you evaluate enterprise agents, JoyAgent's public repo and self-hostable design are worth a pilot — test the DAG engine and DataAgent against a real internal workflow rather than a demo.
  • If you track Chinese AI strategy, watch whether JD's "agent-first" framing (over chat) spreads to other internet incumbents with their own operations to mine.
  • If you build hardware or toys, JD's JoyInside programme is openly recruiting device partners, signalling where a commerce platform thinks embodied AI will first pay off.

Related coverage

Cite this article

NeuroAI (脑机网), "JD open-sourced JoyAgent, a fully open enterprise agent it already runs by the tens of thousands," 14 November 2026, https://neuroai.site/a/na-corp-jd-joyagent

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories