Leaked specs reveal Qualcomm's AI-heavy Snapdragon 8 Gen 3

Smartphone makers hype benchmark scores every single year, but right now, every hardware executive has just one real obsession: can we actually run generative AI directly on a phone?

We might get the real answer much sooner than expected. According to an exclusive leak of internal Qualcomm documents obtained by MSPoweruser, the full specification sheet for the upcoming Snapdragon 8 Gen 3 has surfaced ahead of schedule. Qualcomm is pitching it internally as its debut mobile platform designed specifically for on-device generative AI.

Word on the street is that we could see the flagship silicon debut inside the Samsung Galaxy S24 lineup as early as February next year, and the sheet reveals massive architectural pivots built entirely around local AI workloads.

Rebuilt for INT4 and a 1+5+2 CPU layout

Running heavy models on a pocket device with strict battery limits is a nightmare, so Qualcomm gave its Hexagon NPU an aggressive overhaul.

The leaked specs reveal full precision support across the board, including INT4, INT8, INT16, and FP16. Squeezing down to INT4 precision is crucial because it allows complex generative models to run locally while slashing memory footprint and power consumption.

The Qualcomm Sensing Hub is also getting upgraded with dual micro-NPUs tailored for low-power audio and sensor inputs, backing dual always-sensing cameras that also support INT4 precision.

To feed that kind of processing pipeline, the platform supports up to a massive 24GB of LP-DDR5x memory clocked at 4800 MHz. That memory ceiling is not just for multitasking; it gives large local models enough headroom to breathe.

The CPU cluster sees a notable shift too. Qualcomm dropped older core balancing in favor of a 1+5+2 architecture, packing one prime core at up to 3.3 GHz, five performance cores at 3.2 GHz, and just two efficiency cores at 2.3 GHz. Qualcomm is clearly squeezing efficiency cores to max out sustained multi-core muscle.

Chasing Unreal Engine 5 and Apple's ray tracing

AI is not the only battlefield. After Apple pushed hardware ray tracing on the iPhone 15 Pro, Qualcomm is fighting back on GPU capabilities.

The Adreno GPU in the Gen 3 features hardware-accelerated ray tracing with global illumination, specifically tapping Unreal Engine 5.2 and its Lumen lighting system. Qualcomm claims the chip can even upscale cutscenes up to 8K resolution.

Connectivity is driven by the Snapdragon X75 5G Modem-RF system and the FastConnect 7800 suite, which unlocks Wi-Fi 7 with peak theoretical speeds of 5.8 Gbps alongside Bluetooth 5.4.

Interestingly, the leaked document highlights an initiative dubbed Snapdragon Seamless. Apple and Samsung have locked users into proprietary cross-device systems like AirDrop and Galaxy Connect for years. Qualcomm appears to be building an underlying interoperability layer so any OEM using Snapdragon hardware can create cross-device experiences without reinventing the wheel.

Why it matters

We think Qualcomm's smartest move here is treating on-device AI as an architecture problem rather than a marketing bullet point.

By pushing INT4 support down to the silicon and expanding RAM ceilings to 24GB, the company is preparing smartphones to host local text and image generation models without constantly pinging cloud servers. For daily users, that means generative assistants and on-device photo editing could finally run privately and instantly with zero network latency.

For the industry, it pressures Apple to defend its lead. Apple staked its ground on high-end console gaming, and Qualcomm's answer—pairing Unreal Engine 5.2 Lumen support with serious hardware ray tracing—shows that mobile flagships are going head-to-head on desktop-tier graphics.

That said, shrinking efficiency cores down to just two while ramping five performance cores to 3.2 GHz brings obvious thermal risks. If device makers cannot dissipate that heat, these ambitious local AI and 8K upscaling features will throttle fast. It is definitely worth watching, but buyers should wait for real-world thermals before assuming the performance jump holds up in daily use.