As generative AI continues to rapidly advance, many people want to run large language models on their own computers. But running a 100-billion-parameter model as capable as ChatGPT, Gemini, or Claude requires an enormous amount of memory—even the rumored M5 Ultra with up to 768GB of unified memory this year wouldn’t cut it.
Apple also seems to be aware of this, preparing to once again boost the advantages of Mac’s on-device AI computing. According to Bloomberg According to the latest report from journalist Mark Gurman, Apple’s upcoming M7 Ultra chip could support up to 1.5TB of unified memory, and if this configuration is actually released, Macs would have the opportunity to run trillion-parameter level AI models locally.
Image source:Wccftech
Report: Apple M7 Ultra rumored to support up to 1.5TB of memory, enabling future Macs to run trillion-parameter AI models locally
To understand how impressive 1.5TB is, look no further than the current M3 Ultra.
M3 Ultra offers up to a 32-core CPU, 80-core GPU, 32-core Neural Engine, and 512GB unified memory, with memory bandwidth reaching 819GB/s. 512GB alone is already a rare configuration in current personal computers, and Apple has positioned it as a workstation capable of fully loading large language models with over 600 billion parameters into memory.
The M5 Ultra, expected to launch in late 2026, may feature up to a 36-core CPU, 80-core GPU, and 768GB of unified memory. While the GPU core count hasn’t increased dramatically, the memory capacity is 256GB higher than the M3 Ultra’s maximum configuration—clearly signaling preparation for local AI, 3D rendering, scientific computing, and other professional workloads.
Beyond that is the rumored M7 Ultra, whose maximum memory capacity could double directly to 1.5TB. This is not only about twice that of the M5 Ultra, but also three times the M3 Ultra’s maximum capacity, and even matches the 1.5TB memory option once offered by the 2019 Intel Mac Pro.
The M7 Ultra is designed to support as much as 1.5 terabytes of memory — double the capacity planned for the M5 Ultra — though whether Apple ultimately offers that configuration will depend on the state of the industry. https://t.co/AyZ5GiIRA0
— Mark Gurman (@markgurman) July 12, 2026
The difference is that the Mac Pro back then used standard DDR4 ECC memory, while if the M7 Ultra adopts Apple Silicon’s unified memory architecture, both the CPU and GPU can directly access this 1.5TB of capacity, that would truly be something else.
After the M3 Ultra launched, YouTube channel Dave2D successfully ran DeepSeek R1 with 671 billion parameters on a Mac Studio equipped with 512GB of unified memory.
This model is approximately 404GB in size, requiring about 448GB of GPU memory to run. Since macOS doesn’t grant such a massive amount of memory to the GPU by default, testing also requires manually raising the limit through the terminal. After setup, the M3 Ultra can output approximately 17-18 tokens per second, and the entire Mac Studio consumes less than 200W of power.
If the M7 Ultra truly comes with up to 1.5TB of memory, it should get many local AI enthusiasts excited. Extrapolating from that, if paired with an 8-bit quantized version, the M7 Ultra could theoretically load models with approximately 1.2 trillion parameters.
Using the simplest calculation, 8-bit means each parameter takes up about 1 Byte, so a model with 1.2 trillion parameters would require approximately 1.2TB of space for weights alone, and 1.5TB of unified memory could potentially accommodate it.
That said, even though 1.5TB is tempting, memory shortages remain the main challenge. Bloomberg also reported that while the M7 Ultra was designed with up to 1.5TB as the target, whether Apple will actually offer this configuration will depend on industry conditions at the time. In recent years, AI servers have consumed massive amounts of memory production capacity, making DRAM not only more expensive but high-capacity products increasingly difficult to obtain.
If it actually launches, the price could also be staggering. Before the M3 Ultra unified memory 512GB option was taken down, it was already priced close to $10,000 (without maxing out other specs). Therefore, the top-tier M7 Ultra Mac Studio equipped with 1.5TB unified memory, Wccftech estimates, could reach around $20,000.
Some people might ask, what about the M6 Ultra? It looks like it will be skipped.
Bloomberg previously reported that Apple prioritized upgrading the Neural Engine due to the immense pressure from global AI development. According to sources, Apple’s R&D team completed the design finalization of the next-generation M7 in just six months after finishing the M6 chip design. If they continued to release the M6 Ultra as planned, the architecture could become outdated by then. Therefore, they made a strategic adjustment: after the M5 Ultra, they will skip a generation and directly deploy the M7 Ultra.
Source: KOCPC Chinese