Nvidia has released ARDY code, checkpoints and demos for streaming text- and constraint-controlled 3D motion. Its authors report 33-millisecond generation on an RTX 4090, but that result does not establish end-to-end deployment speed, a 24GB minimum or production economics.
ARDY turns a graphics paper into a runnable motion-generation stack that can respond as a user changes text, paths, keyframes and joint targets. The release gives developers more than curated video, but it remains a research component: its strongest speed result covers generation on one high-end GPU, while integration cost, cleanup and physical execution are unresolved.
ARDY generates motion in segments, conditioning each step on prior motion, the current text prompt and optional spatial goals. Its hybrid representation keeps global root movement explicit while compressing the rest of the body into a latent embedding. A two-stage transformer denoiser predicts the root and then the body conditioned on that root.
That design is intended to preserve direct trajectory control without requiring a fully explicit representation for every body feature. The authors say in the paper abstract that evaluations on HumanML3D and the larger Bones Rigplay dataset showed strong motion quality and constraint adherence. Those are research-team results, not evidence from a shipped game, industrial simulator or robot fleet.
The release repository provides four checkpoints dated July 10, 2026. Two use a 27-joint Core skeleton at 20 frames per second, with generation horizons of 40 or eight frames. Two use the 34-joint Unitree G1 skeleton at 25 frames per second, with horizons of 52 or eight frames. The frame rates describe motion output, not measured inference throughput.
Users can change text during playback, set root waypoints, steer with a keyboard, or impose full-body and sparse hand or foot constraints. The browser demo can export and reload sessions. A command-line path writes NPZ motion files and, for the G1 skeleton, MuJoCo qpos CSV files.
The paper frames ARDY as a research system for animation, simulation and humanoid robotics. The relevant commercial question is whether Nvidia's GPU-centered implementation saves enough animation work to justify the rest of the deployment stack.
The authors report in the full primary paper that their interactive demo ran on an RTX 4090 with average generation latency of 33 milliseconds for a four-step diffusion model and 63 milliseconds for a 10-step version, which they say offers slightly better control accuracy. Both generated two-second windows at 20 frames per second. The slower version used one already-generated buffer frame to conceal latency while new motion was computed.
That is meaningful evidence of interactive generation, but its scope is narrow. The paper calls it generation latency; it does not present the figure as an end-to-end measurement that includes prompt encoding, post-processing, scene logic, rendering or an engine's other work. The repository separately warns that the buffer must be sized so generation finishes before the buffered playback frames run out.
The hardware claim is narrower too. The repository says its main test configuration was Ubuntu 22.04 with an RTX 4090, while the model card names both an A100 and RTX 4090 as test hardware and lists Ampere, Hopper and Blackwell compatibility on Linux. Nvidia's reference specifications give the RTX 4090 24GB of GDDR6X memory, 450 watts of total graphics power and an 850-watt system-power recommendation. None of these sources defines 24GB as ARDY's minimum or demonstrates equivalent support on an arbitrary 24GB GPU.
Text encoding adds its own resource choice. ARDY uses LLM2Vec on top of Meta-Llama-3-8B-Instruct. Developers must obtain access to that gated Meta model and provide a Hugging Face token. The default GPU bfloat16 text-encoder mode uses about 14GB of VRAM; CPU modes use less GPU memory but make prompt encoding slower.
Real-time, text-driven motion predates ARDY. The 2024 DartControl paper, written by ARDY co-authors Kaifeng Zhao and Siyu Tang with Gen Li, already described sequential, real-time motion generation from changing natural-language descriptions. It handled spatial control through latent-noise optimization or reinforcement learning.
ARDY's more defensible distinction is the combination: one streaming generator is directly conditioned on online text and flexible kinematic constraints, including targets beyond the immediate generation window. In the ARDY authors' own comparison, DartControl needs optimization or an additional control policy for spatial goals; MotionStreamer supports online text and variable history but not kinematic goals; and DiP combines text and joint goals but with shorter history and prediction horizons. That is a broader native control interface, not the invention of interactive generative animation.
The quantitative comparison also needs its protocol. The authors tested ARDY and DiP on nine-second HumanML3D sequences seeded with one second of ground-truth motion, using goals either inside the current generation window or at the end of the sequence. They report that ARDY surpassed DiP in both cases. The authors designed and ran that test; it is not independent validation, and it does not measure engine integration or cleanup time.
Nvidia also positions ARDY beside, rather than above, its own substitutes. Kimodo targets controllable offline authoring, MotionBricks targets fast motion in-betweening, and ARDY output can be passed to GEAR SONIC for physical robot tracking. The handoff is important: ARDY produces kinematic poses, not a proof that a robot can execute them safely or that a character will react correctly to a populated scene.
The codebase uses Apache-2.0, but the checkpoints and data do not inherit that license. The Core Horizon40 model card says the checkpoint is ready for commercial or non-commercial use under the Nvidia Open Model Agreement. The repository tells users to review separate terms for checkpoints, datasets and third-party software.
The setup is also specific: Python 3.10 or newer, PyTorch 2.4 or newer with a matching CUDA build, Linux, and build tools for a bundled C++17 motion-correction extension. Optional TensorRT acceleration requires a CUDA 12-capable Nvidia driver and access to Nvidia's package index. This is locally runnable, but not hardware-neutral or one-click.
No retained source provides a hosted inference price, engine plug-in, deployment-cost comparison or production support commitment. A developer-focused briefing likewise identifies production-engine integration cost as unresolved.
The Core Horizon40 model card lists 326 million parameters, 630 hours of motion capture for training and a 70-hour held-out test split. That scale helps explain why ARDY can cover locomotion, gestures, combat, dancing and everyday activities, but Nvidia's own documentation sets clear boundaries.
Generated motion can contain foot skating or jitter, and it may fail to follow a text prompt. Each checkpoint outputs one skeleton. The model is intended for realistic human movement, not cartoon or deliberately nonphysical motion, and it has no awareness of nearby objects. Optional post-processing can reduce foot skating and improve constraint following, but it is slower and disabled by default in the interactive demo.
The full paper describes ARDY as purely kinematic and lacking physical dynamics. A developer may therefore gain live text and path control while still needing collision logic, retargeting, scene interaction, a physics controller and manual or automated cleanup. Nvidia's model card also says use-case-specific testing is required before safe and effective deployment.
The immediate test for a studio or robotics team is not whether the demo moves at interactive speed. It is whether ARDY lowers total authoring and runtime cost once text encoding, motion correction, retargeting, physics, scene logic, rendering and failure recovery are included.
That requires end-to-end latency and memory benchmarks across more than one GPU, including smaller-memory setups that move text encoding to the CPU. It also requires like-for-like trials against motion matching, animation state systems and narrower generative tools, measured on usable output, cleanup time and recovery from failed prompts or constraints.
The repository says a Rigplay-trained checkpoint for the SOMA skeleton is coming, without a date. Until broader skeleton support, independent production tests and full-pipeline economics arrive, ARDY is best treated as a substantial research release with a strong native control proposition—not yet a replacement for the systems that make generated motion deployable.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
Nebius has arranged its first senior secured facility against an operating GPU deployment and one customer’s cash flows. The deal adds project-level debt to its expansion toolkit, but an unnamed customer and a larger rival financing limit what it proves about the rest of Nebius’s backlog.
Huawei publicly displayed a 16-cabinet Atlas 950 configuration rated at 1 EFLOPS in FP8, providing tangible evidence of its system-scale AI strategy while leaving price, power use and sustained workload performance undisclosed ahead of the full system's planned fourth-quarter release.
UK testing places leading open-weight models four to seven months behind selected closed-model cyber results, yet longer attack chains, U.S. benchmarks and mixed cost comparisons show why that interval is a warning signal rather than a universal capability clock.
Moonshot AI's Kimi K3 added to a global technology selloff with near-frontier performance, but unreleased weights, mixed cost comparisons and a recommended 64-accelerator deployment leave its effect on chip demand unresolved.
Two binding EU decisions require Google to give rival AI services comparable access to 11 Android features and offer eligible search competitors a restricted, anonymized dataset, but phased deadlines, certification, pricing and privacy safeguards leave the competitive effect unproven.
Databricks has signed a term sheet for a Coatue-led financing at a $188 billion valuation, while unidentified sources put the round at $3 billion. The proposed capital would deepen its push into AI governance, data agents and operational databases, but the transaction remains open and the company supplied no new operating figures.
OpenAI says a handful of GPT-5.6 Sol file-deletion reports most commonly involved Full Access without sandboxing or Auto-review. Its own evaluations show a more complicated risk picture, and the company has not yet published an incident rate or evidence that its promised safeguards stop the failure.
Netflix says generative AI workflows were used on roughly 300 titles in 2026, but its only quantified example tied to that disclosure covers 17 minutes and does not establish how much finished material, spending or labor changed across the slate.
A rogue browser extension can trigger Claude for Chrome’s fixed Gmail, Docs, Calendar and business workflows; the attack is constrained by default approvals but can run silently for users who enabled unattended action.
Gemini 3.5 Pro missed its expected June rollout. An anonymous-source account says a late-June data change fell short of Google's coding goals, but Google has confirmed only partner testing—not the reported cause, a new date, or public results and pricing.
Gold Eagle has begun collecting and prioritizing AI-discovered software vulnerabilities, but the voluntary federal clearinghouse has not disclosed results, operating rules or the resources that would turn findings into deployed fixes.
Google is bringing Gemini Omni editing and a reusable face-and-voice avatar to Vids, but the sharper distinction is account-level identity across Vids and Gemini rather than a new category of AI video; Vids already offered Veo generation and customizable avatars, while specialist rivals already sell digital presenters.