In March 2024 Qualcomm dropped the Snapdragon 8s Gen 3 into phones that cost half what a Galaxy S24 Ultra did, and those phones could still run Google’s Gemini Nano and a 7-billion-parameter Llama model without ever touching a server. That single fact is the whole story of this chip, and looking back from 2026, I think almost every review at the time buried it under gaming benchmarks. I made the opposite case the day it launched, and two years on I’d make it harder.
The flagship got the spotlight, and on paper it earned some of it. The 8 Gen 3 shipped the previous October on TSMC’s N4P process with a Cortex-X4 prime core at 3.3 GHz over a 1+3+2+2 Kryo cluster, 12 MB of L3, and the Adreno 750, which landed about 25% faster than the 740 with a real ray-tracing jump and global illumination that showed up in the three or four games that actually supported it. It ran roughly thirty percent more CPU and twenty percent better efficiency than the Gen 2, clean generational math that left me cold.
What made me sit up was the Hexagon NPU. Qualcomm claimed it ran 98% faster than the 8 Gen 2’s, and the demo that mattered wasn’t a TOPS slide, it was a 7-billion-parameter Llama 2 generating text at roughly 20 tokens a second entirely on the phone, weights streaming out of LPDDR5X. Twenty tokens a second isn’t fast in any absolute sense. It’s faster than you read, though, and that turns out to be the only threshold that matters when you’re watching a reply fill in on a screen in your hand. I dug into exactly where that performance wall sits, and why it’s the NPU and not the CPU that decides it, over in my piece on running local LLMs on Android.
Then March came and Qualcomm did the thing that actually moved the needle. The 8s Gen 3 wasn’t a downclocked flagship, and it wasn’t a rebadged 7-series either, which is what most of us expected. It was its own part: one Cortex-X4 at 3.0 GHz, four A720 cores at 2.8, three A520s at 2.0, a 1+4+3 layout that doesn’t match the flagship’s at all. The GPU dropped to the Adreno 735, it lost global illumination, and the modem was the older X70 pulled up from the 8 Gen 2 generation. On a spec sheet it reads like a compromise, and in raw gaming it is one.
But Qualcomm kept the piece that counted. The 8s Gen 3 carried the flagship’s entire on-device generative-AI stack, Qualcomm’s AI Hub included, and ran Baichuan-7B, Llama 2, and Gemini Nano locally. That was the whole point of the thing. The flagship proved a phone could hold a language model in memory and talk back; the 8s Gen 3 made that capability cheap enough to put in a Redmi Turbo 3 or a Poco F6, phones that undercut a Galaxy flagship by half. Honor, iQOO, Realme, and Xiaomi’s sub-brands all lined up for it within weeks. This is the moment on-device AI stopped being a flagship party trick and started being a baseline expectation for anything above the true budget tier, and I don’t think it gets enough credit for that.
The competition tells you why it worked. Apple’s A17 Pro, over on a 3nm node, beat the 8 Gen 3 on raw single-core and GPU by a comfortable margin, and it wasn’t close on peak silicon. What Apple didn’t have in early 2024 was any real on-device LLM story; the Neural Engine was strong at the camera-and-Siri workloads it was tuned for, but Apple wasn’t handing developers a way to run a 7B model on the phone. MediaTek’s Dimensity 9300 was the genuine Android rival on the flagship end, and a good one, but Qualcomm had the thing MediaTek couldn’t match that spring: an ‘s’ part that pushed the same generative-AI capability down a full price tier while MediaTek was still fighting for the top. Fragmentation, the thing everyone treats as Android’s curse, was Qualcomm’s weapon here.
I’m not going to relitigate the sustained-performance argument in this post, because the 8 Gen 3’s thermal throttling under long loads is its own can of worms and deserves its own teardown. What I’ll say is that the AI workloads people actually ran on these chips were bursty, not sustained, so the throttling that hammered gaming benchmarks mattered a lot less for the use case that turned out to define the generation.
Two years and two full generations on, with the 8 Elite and the 8 Gen 5 now shipping, the 8s Gen 3 is the one I’d point to as the chip that aged best, and not for anything on a benchmark chart. It made a language model on mid-priced hardware feel ordinary, and that shift has since hardened into table stakes across the 2026 flagships and a lot of what sits under them. For the longer arc of how Qualcomm got here, I traced it in two decades of Snapdragon. The company spent its keynote selling the hero silicon while the part that quietly rewired what a cheap phone can do shipped without a poster, and I’m not convinced the industry has fully clocked that even now.