Apple M6: Dual Neural Engine, memory speeds up to 170 GB/s, and what this means for local AI

On August 25, 2026, Apple unveiled the M6 chip, its first in-house chip manufactured using a 2-nanometer process, highlighting its performance for on-device artificial intelligence. Key new features include the “Dual 16-core Neural Engine” and higher memory bandwidth. What is truly new here, and what does it mean for people who want to run language models locally?

What Apple Announced

The M6 features a 12-core CPU, a 12-core GPU with neural accelerators in each core, and a dual-neural engine. According to Apple, it uses unified memory of up to 32 GB with a bandwidth of up to 170 GB/s, which is approximately 10% more than its predecessor, the M5. The first computer to feature this chip is the new Mac mini, which went on sale on September 22, 2026. Along with the M6, Apple also introduced the M5 Ultra chip with a quad-chip configuration.

Dual Neural Engine: What's Confirmed

Previous M-series chips had a single 16-core neural engine. The M6 has two, and according to Apple, system libraries can automatically utilize both simultaneously, allowing apps to run AI models faster without changing the code. Apple reports up to twice the maximum computational performance of the neural engine compared to the previous generation. This is a peak value under ideal conditions, not a guaranteed speedup for every task.

What does “30%” mean?

Some descriptions mention a “30% increase in AI performance.” According to Apple, this refers to a nearly 30 percent increase in the graphics processor’s peak computational performance for AI tasks compared to the M5, which is intended to improve task processing speed for on-device language models. The neural engine sees an increase of up to twofold. These two figures should not be confused.

Why Memory Is Important

When generating text with a language model, large amounts of data are read from memory at every step. Therefore, the speed of local models tends to depend on memory bandwidth, not just the number of compute cores. This is a general principle, not a claim from a single source. 170 GB/s represents a slight improvement over the M5 (153.6 GB/s), but it is significantly less than that of the M5 Pro and M5 Max chips. The M5 Pro offers up to 307 GB/s, and the M5 Max offers up to 460 to 614 GB/s, depending on the configuration. The base M6 is therefore not designed for the largest models.

Limit: maximum 32 GB

The M6 supports up to 32 GB of unified memory, so larger models won’t fit on it. Anyone who needs more memory will have to opt for more powerful chips.

Privacy and Local AI

On-device processing means that data for tasks running locally does not have to leave the computer. This is an advantage for privacy and for working offline. However, this is not guaranteed for every feature: what happens locally and what happens in the cloud depends on the specific application and service. This is our observation, not a manufacturer’s claim.

Practical examples and expectations

• Developers. Apple claims that the M6 speeds up compilation, file indexing, and running multiple simulators in Xcode. We haven’t found any specific independent benchmarks yet.
• Local language models. Faster prompt processing and higher memory throughput should help with smaller models. Actual speed—such as the number of tokens per second—will only be determined by testing specific models.
• Video and other tasks. We did not find any data in the sources regarding the acceleration of video editing thanks to the neural engine.

What we haven't confirmed

• Vibe Coding and “macOS 27 moves AI from toys to tools.” We have not confirmed details about this feature in relation to the M6.
• “A radical performance boost for agent-based AI.” Apple cites peak performance figures, not real-world results.
• Availability and prices in Slovakia. We do not have data for our market.

Conclusion

The M6 is a logical step toward local artificial intelligence: two neural engines, more powerful graphics cores, and slightly higher memory bandwidth. Independent tests will reveal the exact benefits for the average user, especially with local language models, where the 32 GB memory limit also plays a role. When deciding on a purchase, it’s therefore worth focusing on specific tasks rather than the top-of-the-line numbers from the presentation.