Lenovo puts a 120-billion-parameter model on a laptop with Nvidia's RTX Spark
Lenovo unveiled the first laptops built on Nvidia's RTX Spark platform at IFA 2026 in Berlin, and a Chinese robotics lab showed how to keep training a policy while the robot is still working — two stories that point at the same shift: AI is moving off the cloud and onto the box under your hand.
Lenovo and Nvidia put a 120-billion-parameter model inside a laptop. At IFA 2026, Lenovo announced the Yoga Pro 9n and the Yoga 9n 2-in-1 — the first notebooks built on Nvidia's RTX Spark silicon, a system that pairs a 20-core Grace Arm CPU with a Blackwell RTX GPU carrying up to 6,144 CUDA cores over an NVLink-C2C link, delivering 1 PFLOPS of AI compute in FP4. The flagship Pro 9n ships with up to 128GB of 9,400 MT/s unified memory that the CPU and GPU share, which is the number that actually matters: a 120-billion-parameter model with a 1M-token context can sit fully resident on the device. The 2-in-1 caps memory at 64GB but adds a 360-degree hinge, dual-surface pen support, and a haptic Force Pad. Nvidia's stack lets the laptops coordinate multiple local agents across apps for creative work without round-tripping the cloud, and Lenovo's SVP Jun Ouyang framed the line as a creator-developer tool rather than a productivity PC.
Pricing and exact ship dates weren't disclosed — Nvidia previously said the first wave (Lenovo, Asus, Dell, HP, Microsoft, MSI) lands this fall, with Acer and Gigabyte to follow. Expect enthusiast pricing, well above the Snapdragon X2 Copilot+ class. The interesting thing isn't the laptop, it's the category: a year ago "AI PC" meant an NPU doing 40 TOPS of background tricks; the RTX Spark machines are the first generation where "AI PC" describes a real local compute platform, not a marketing sticker. RTX Spark earlier this year was a desktop workstation — putting the same silicon in a 3.6-pound Yoga is the step that makes "local 100B+ model" a category rather than a demo.
Astribot's SmoothRL keeps a robot learning while it keeps moving. Beijing-based Astribot released SmoothRL, an online reinforcement-learning framework designed for the messy reality that a robot doesn't stop and wait for a gradient step. Most online-RL setups force the model to pause while a new policy is computed; SmoothRL aligns the learning loop with the robot's actual execution timeline, splitting each action chunk into committed, in-flight, and discarded regions and only updating gradients over the parts that actually ran. A temporal-value critic carries the full action sequence into the loss, and a trajectory-continuity constraint stops the policy from jerking between updates. The paper, on arXiv as 2608.29768, reports steady success-rate gains on three real-robot tasks — dynamic tossing, pen capping, and box opening — across a single continuous online fine-tuning run, with two human-intervention modes (VR teleop for precision, joystick for residual corrections) that feed straight into the learning loop. The pitch is that a deployed robot can keep getting better at its job without taking the job offline, which is the precondition for the "continual learning" agenda Richard Sutton and others have been pushing for years.
China publishes a package of national standards for large AI models. A set of new Chinese national standards for large AI models has been formally released, codifying safety and quality requirements across the training-data, fine-tuning, evaluation, and deployment stages of the model lifecycle. The package is part of the broader six-document AI policy stack Beijing published this month, and follows the country's first national large-model standard from earlier this year. The practical effect: Chinese cloud and model providers now have a single compliance checklist to point at when procurement officers, regulators, and overseas customers ask what "safe" means in their stack. The interesting question is whether the standards become a de facto template for the rest of the Global South — the same way the earlier GenAI safety standard quickly became reference reading for ASEAN and African Union regulators.
What to watch: Lenovo and Nvidia have yet to name a ship date or a price on the Yoga Pro 9n — the more the silence drags past October, the more it suggests a "creator workstation" price tag, not a mainstream one.
If RTX Spark laptops ship in volume under $3,000, the case for cloud-only AI agents on consumer hardware gets a lot weaker. If they don't, local 100B-parameter models stay a developer and prosumer niche for another year. Tell us in the comments.
Sources: Lenovo + Nvidia IFA 2026 announcement (Thurrott) · Yoga Pro 9n hands-on (How-To Geek) · 联想携手英伟达发布新款AI PC (PChome) · Nvidia — RTX Spark IFA 2026 · Astribot — SmoothRL project page · SmoothRL on arXiv (2608.29768) · 星尘发布在线强化学习框架SmoothRL (Sina Finance) · China large-model standards package (Aibase)
