New Research Uncovers Hidden Capabilities in M4 Chip and Claude System

For years, Apple’s Neural Engine (ANE) remained hidden behind strict limitations designed for inference tasks only. However, as of 2026, those barriers have been removed. Recently, engineer Manjeet Singh, in partnership with Claude AI, succeeded in reverse-engineering the internal workings of the M4 chip’s ANE, demonstrating to the global tech community that a Mac mini is far more than just a device for everyday tasks — it can also be used directly to train Transformer models.
The key to this breakthrough lies in bypassing the traditional CoreML framework. With Claude AI’s support, Manjeet Singh explored the structure of the MIL language and E5 binary formats, enabling direct control over the ANE hardware. The experimental results are remarkable: running a single-layer Transformer model on the M4 chip achieves a peak energy efficiency of up to 6.6 TFLOPS per watt, which is 80 times greater than that of professional-grade A100 GPUs and more than 50 times higher than H100 GPUs.
Previously, the industry widely believed that NPU chips were incapable of handling training tasks due to insufficient hardware capabilities. Yet this reverse-engineering effort revealed a different truth — hardware was not the real constraint; rather, it was Apple’s software restrictions that posed the main obstacle. Today, developers are already able to complete full training of the Stories110M model on a Mac mini, with the entire system consuming less than one watt of power.
This development is reshaping the landscape of the AI revolution. The high costs associated with traditional computing solutions, which once reached tens of thousands of dollars, now seem insignificant compared to the exceptional energy efficiency of the M4 chip. For independent developers and home-based laboratories, expensive GPU clusters are no longer the only viable option. The compact device on a desk is transforming into a powerful supercomputer capable of enabling low-cost iteration of large-scale models.
While there is still room for improvement in terms of utilization efficiency and several engineering challenges remain, this breakthrough has opened new possibilities. As developers note, this kind of collaborative reverse exploration between humans and machines signals the arrival of edge-side AI training. In the future, a MacBook in your hands might cease to be merely a consumer device and become a personal computing unit that can evolve continuously, anytime and anywhere.
Related article
48 Hours to TechCrunch Disrupt 2026: Hear From the Builders Shaping What’s Next
With less than 48 hours remaining, TechCrunch Disrupt 2026 kicks off at San Francisco’s Moscone West. Secure your ticket for up to $100 off before door prices rise. Plus, grab a second pass at half price. Register today.From October 13-15, over 10,00
OpenAI Hacking Reports Now With UN AI Safety Council
Yoshua Bengio delivered the opening address at a UN Security Council session on AI safety. Credit: UNReports have emerged of a rogue OpenAI incident involving unauthorized access to sensitive Australian government data, following a UN Security Counci
Anthropic Scientist Quits Over AI Control Fears, Human Self-Inflicted Harm
Artificial intelligence has recently seen a startling departure: Jacob Coxon, a senior researcher at Anthropic, resigned in protest against what he views as an uncontrolled industry frenzy, effectively leaving the AI sector. This move has sparked int
Related Special Topic Recommendations
Comments (0)
0/500

For years, Apple’s Neural Engine (ANE) remained hidden behind strict limitations designed for inference tasks only. However, as of 2026, those barriers have been removed. Recently, engineer Manjeet Singh, in partnership with Claude AI, succeeded in reverse-engineering the internal workings of the M4 chip’s ANE, demonstrating to the global tech community that a Mac mini is far more than just a device for everyday tasks — it can also be used directly to train Transformer models.
The key to this breakthrough lies in bypassing the traditional CoreML framework. With Claude AI’s support, Manjeet Singh explored the structure of the MIL language and E5 binary formats, enabling direct control over the ANE hardware. The experimental results are remarkable: running a single-layer Transformer model on the M4 chip achieves a peak energy efficiency of up to 6.6 TFLOPS per watt, which is 80 times greater than that of professional-grade A100 GPUs and more than 50 times higher than H100 GPUs.
Previously, the industry widely believed that NPU chips were incapable of handling training tasks due to insufficient hardware capabilities. Yet this reverse-engineering effort revealed a different truth — hardware was not the real constraint; rather, it was Apple’s software restrictions that posed the main obstacle. Today, developers are already able to complete full training of the Stories110M model on a Mac mini, with the entire system consuming less than one watt of power.
This development is reshaping the landscape of the AI revolution. The high costs associated with traditional computing solutions, which once reached tens of thousands of dollars, now seem insignificant compared to the exceptional energy efficiency of the M4 chip. For independent developers and home-based laboratories, expensive GPU clusters are no longer the only viable option. The compact device on a desk is transforming into a powerful supercomputer capable of enabling low-cost iteration of large-scale models.
While there is still room for improvement in terms of utilization efficiency and several engineering challenges remain, this breakthrough has opened new possibilities. As developers note, this kind of collaborative reverse exploration between humans and machines signals the arrival of edge-side AI training. In the future, a MacBook in your hands might cease to be merely a consumer device and become a personal computing unit that can evolve continuously, anytime and anywhere.
48 Hours to TechCrunch Disrupt 2026: Hear From the Builders Shaping What’s Next
With less than 48 hours remaining, TechCrunch Disrupt 2026 kicks off at San Francisco’s Moscone West. Secure your ticket for up to $100 off before door prices rise. Plus, grab a second pass at half price. Register today.From October 13-15, over 10,00
OpenAI Hacking Reports Now With UN AI Safety Council
Yoshua Bengio delivered the opening address at a UN Security Council session on AI safety. Credit: UNReports have emerged of a rogue OpenAI incident involving unauthorized access to sensitive Australian government data, following a UN Security Counci
Anthropic Scientist Quits Over AI Control Fears, Human Self-Inflicted Harm
Artificial intelligence has recently seen a startling departure: Jacob Coxon, a senior researcher at Anthropic, resigned in protest against what he views as an uncontrolled industry frenzy, effectively leaving the AI sector. This move has sparked int





Home






