Four hardware announcements from November 10, all about getting AI and new physics out of the lab and into things that move or sit at the edge.
MemryX MX3+, ten times the MX3
MemryX unveiled the MX3+, an edge AI accelerator with about ten times the peak compute and nearly ten times the on-chip memory of its MX3, inside the same low power envelope. The company keeps its multi-chip approach. Cards can stack from one chip up to sixteen, and a 16-chip setup is meant to pass 1 POPS of peak AI compute with 2.16 GB of on-chip memory spread across the chips. The compiler splits a network across the accelerators so the host sees one device.
The target workloads say a lot about where edge AI has gone. MX3 was mostly about vision. MX3+ is pitched at transformers, small language models and vision-language-action models for robots, alongside higher-resolution vision. Engineering samples are due in late 2027 and volume production in mid-2028, so this is a roadmap, not a product. MemryX hasn’t given per-chip TOPS or power figures yet. The next chip, MX4, is aimed at multimodal and large language model inference.
Tianji’s 21-DoF hand and NVIDIA’s GR00T 2
At CoRL 2026 in Austin, Tianji Intelligence showed a 21-degree-of-freedom hand under its GENTO brand: five active joints in the thumb and four in each finger. It weighs 750 g without cables, grips objects up to 100 mm across with 10 N of fingertip force, repeats positions to within ±0.05 mm and is rated for a million no-load grasps.
The more interesting part is who uses its hardware. Tianji built Marvin Pro, a dual-arm platform, for NVIDIA. In his CoRL keynote, NVIDIA’s Jim Fan showed GR00T 2, a model that predicts future video frames and robot actions together. Human hand motions were captured with a wearable exoskeleton, then GR00T 2 ran on a Marvin Pro with two dexterous hands to do tasks such as making banana yogurt. Learning from human demonstration needs hands close to human ones, and that is where the 21 joints matter.
Appear moves solid-state cells to pre-production
San Francisco’s Appear Inc. says it has finished field and manufacturing trials of its solid-state battery and started pre-production with a manufacturing partner it won’t name. Limited commercial output is planned for mid-2027 and full scale in 2028. The first targets are motorcycles, e-bikes and scooters. Appear published no energy density, cycle life or chemistry, so treat the claims with care until a spec sheet appears. Two-wheelers are a sensible place to start, though. Packs are small, safety matters a lot, and buyers will pay more for a battery that won’t catch fire in a hallway.
Special Optics builds objectives for quantum labs
Special Optics, part of AMETEK through Navitar, launched Q-Series microscope objectives made for neutral-atom and cold-atom quantum work. They’re color-corrected across several laser wavelengths and handle high laser power, so labs can trap and manipulate more atoms at once, with support for rubidium, strontium, barium and ytterbium. Until now many groups adapted catalog optics or paid for custom designs. An off-the-shelf part is a small sign that neutral-atom quantum computing is turning into an industry with a supply chain.