192GB. That's the unified memory capacity reportedly coming in AMD's next Halo-series processor, and for anyone running large language models locally, it's a meaningful jump.
Leaked specs point to a chip codenamed "Gorgon Halo" - likely the Ryzen AI Max+ 495 - with 192GB of unified memory, up from the 128GB ceiling on the current Ryzen AI Max 395. "Unified memory" means the CPU and GPU share one memory pool, so you can run AI models on the integrated graphics without needing a separate dedicated card. That architecture is what makes the Halo line attractive for local AI work in laptops and compact desktop systems.
What 192GB Actually Gets You
Memory is the hard limit on which models you can run locally. A 70-billion parameter model - roughly what you need to approach mid-tier GPT-4 performance on reasoning tasks - requires around 40GB just to load at reduced precision (quantization cuts the number's size without eliminating the parameter count). At 128GB, the current Max 395 can handle those models with room for context, but it gets tight with the larger 100B+ class models coming out of Meta, Mistral, and Google.
At 192GB, you'd have enough to run a 70B model at full 16-bit precision with memory to spare, or load some of the larger mixture-of-experts architectures that currently require server hardware or cloud compute.
The Price Problem
The catch is cost. The memory in these chips is HBM - high-bandwidth memory - which is both expensive to produce and currently in short supply. Laptops with the existing Ryzen AI Max 395 already run $3,000 to $5,000+. Adding 50% more HBM to the die will push prices higher, possibly substantially.
No official release date has been announced; this is leak-level information, not a confirmed product. AMD's current roadmap suggests a 2026 release is plausible. Speculation about a follow-up chip reaching 256GB in 2027 is circulating but has no confirmed basis.
For the local LLM community, the trajectory is clear: consumer-grade hardware is closing in on the memory specs that serious model hosting requires. Whether the price comes down fast enough to matter for individual users is the open question.