AMD Ryzen AI Max+ PRO 495 Leaks 192GB

// 46d agoNEWS

AMD Ryzen AI Max+ PRO 495 Leaks 192GB

A leaked PassMark entry suggests AMD's next Gorgon Halo APU could hit 192GB of unified memory, up from the current 128GB ceiling on Ryzen AI Max+. That's a meaningful jump for local AI workloads, but it's still a leak, not an official launch.

// ANALYSIS

This is less about raw compute and more about memory headroom. For local LLM users, 192GB is the kind of spec that makes bigger quantized models and longer contexts feel practical instead of theatrical.

–AMD's official CES 2026 materials say Ryzen AI Max+ topped out at 128GB unified memory, so 192GB would be the first real capacity bump in this family.
–The leak appears tied to an HP validation system with 188GB usable memory, which is exactly the sort of footprint you see in preproduction OEM samples.
–The "192GB VRAM" framing is misleading: this is shared unified memory, not discrete GPU VRAM, so bandwidth and OS overhead still matter.
–If Medusa Halo really moves to LPDDR6 in 2027-28, 256GB becomes plausible in premium workstation-class machines, but pricing will decide whether anyone actually ships it.
–For the local AI crowd, the bigger takeaway is that AMD is turning memory capacity into a differentiator, not just chasing TOPS.

// TAGS

ryzen-ai-max-pro-495llminferencegpuedge-ailocal-first

DISCOVERED

46d ago

2026-05-04

PUBLISHED

46d ago

2026-05-04

RELEVANCE

9/ 10

AUTHOR

PromptInjection_

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL38m ago

Z.ai has released GLM-5.2, a flagship open-weights Mixture-of-Experts model designed for long-context software engineering and complex agentic tasks.

Z.ai (formerly Zhipu AI) has launched GLM-5.2, a flagship Mixture-of-Experts (MoE) large language model under an MIT open-source license. The model boasts 744 billion total parameters (40 billion active per inference call) and supports a native 1-million-token context window designed specifically to handle complex coding and agentic workflows. To optimize efficiency, GLM-5.2 introduces "IndexShare" to reuse indexers across sparse attention layers—slashing per-token FLOPs by 2.9×—and incorporates an improved Multi-Token Prediction (MTP) layer for faster speculative decoding. Developers can run the model locally using quantized weights or access it via API platforms.

VIDEO1h ago

Anthropic details Claude Code steering methods

Anthropic has detailed seven steering methods to customize and optimize context in its command-line developer tool, Claude Code. These mechanisms, which include path-scoped rules, custom skills, and subagents, help developers manage context bloat and improve code generation accuracy.

UPDATE1h ago

ElevenLabs Music v2 showcases multilingual synthesis

Matt from ElevenLabs demonstrated the cross-lingual capabilities of ElevenLabs Music v2, showcasing tracks that blend genres like Britpop and Samba with foreign languages. The demonstration highlights the model's ability to maintain musical style coherence while singing fluently across multiple target languages.

AMD Ryzen AI Max+ PRO 495 Leaks 192GB