Two major technology companies are making it dramatically easier to run powerful AI models directly on your personal computer, without sending your data to the cloud. AMD announced Day 0 support for Qwen 3.8 27B, a 27-billion-parameter language model, on consumer-grade hardware, while Google unveiled HEIR, an open-source compiler that lets AI models process encrypted data. Together, these developments suggest the industry is moving toward a future where privacy and capability are no longer competing priorities.

What Does Day 0 Support Actually Mean for Developers?

When AMD says it's delivering "Day 0 support," it means developers can start using Qwen 3.8 27B the moment the model becomes publicly available, without waiting for hardware manufacturers or software vendors to optimize their systems. This is a significant shift in how the AI ecosystem typically works. Historically, new models arrive, and then months pass before they run efficiently on consumer hardware. AMD is eliminating that lag.

The company tested Qwen 3.8 27B on two types of consumer-accessible hardware. On an AMD Ryzen AI Max+ processor, the model achieved 24.5 tokens per second, meaning it can generate roughly 24 words of text every second. On a single AMD Radeon AI PRO R9700 graphics card with 32 gigabytes of memory, performance jumped to 51.8 tokens per second. For context, tokens are small units of text; most people read at roughly 200 to 300 words per minute, so 24 to 51 tokens per second is fast enough for real-time conversation.

The model requires at least 24 gigabytes of graphics memory to run comfortably, which narrows the audience to users with higher-end consumer hardware or professional workstations. However, AMD is positioning this as a path for developers and power users to experiment with state-of-the-art models without relying on cloud services or paying per-query fees.

How to Run Advanced AI Models Locally on Your PC

• Use LM Studio: A graphical interface that lets non-technical users download and run Qwen 3.8 27B directly on supported AMD systems, with no coding required. Users can test prompts and workflows immediately.
• Integrate with Lemonade: AMD's local-first developer platform simplifies deployment by handling hardware-aware optimization across CPU, GPU, and NPU resources, allowing developers to package AI inference alongside their applications.
• Leverage llama.cpp: An open-source inference framework that powers the underlying performance benchmarks; developers can use this directly for custom implementations on AMD hardware.

AMD designed Lemonade specifically to solve a real developer problem: building AI applications that work across different hardware configurations. Instead of writing custom code for each processor type, developers can use Lemonade as a lightweight inference layer that communicates with the model through standard API patterns. This dramatically reduces complexity and time-to-market.

Why Google's Encrypted AI Approach Changes the Privacy Equation

While AMD focuses on making local inference accessible, Google is tackling a different problem: how to let cloud services process sensitive data without exposing it. The company introduced HEIR, an open-source compiler that converts pre-trained AI models to operate on encrypted data. This means a cloud service could provide personalized recommendations, detect fraud, or identify security threats without ever seeing the actual user data.

Homomorphic encryption, the cryptographic technique underlying HEIR, has historically been too slow for practical use. Google's compiler project is designed to change that. The company demonstrated four real-world applications compiled with HEIR, including a deep learning recommendation model, credit card fraud detection, network intrusion detection, and a hotword detector for voice-activated AI assistants. Each application was tested on a single-threaded CPU, showing that performance is improving rapidly.

"Homomorphic encryption is another powerful tool we're adding to our private computing toolkit. Like private information retrieval, and in contrast to hardware-based solutions, homomorphic encryption's strong security and privacy guarantees are purely cryptographic,"

Google's research team, as stated in the HEIR announcement

The practical implication is significant: industries like healthcare and finance, which face strict regulations around data sharing, could now use cloud-based AI services without violating privacy laws. A hospital could send encrypted patient data to a cloud service for diagnostic analysis, and the service provider would never see the underlying information.

What's Driving This Shift Away From Cloud-Only AI?

The convergence of these two approaches reflects broader industry concerns. On one hand, running AI locally eliminates latency, reduces cloud costs, and keeps proprietary data off third-party servers. On the other hand, encrypted inference allows organizations to benefit from powerful cloud services while maintaining cryptographic privacy guarantees. Neither approach is perfect for every use case, but together they offer developers and enterprises more options than existed even a year ago.

Google has already partnered with hardware accelerator companies including Belfort, Niobium, Cornami, and Optalysys to speed up homomorphic encryption. The company is also collaborating with academic institutions including Georgia Tech, Carnegie Mellon, UC Santa Barbara, and Purdue. Four peer-reviewed publications have already been built on HEIR, with more in preparation, suggesting the technology is moving from research into practical deployment.

For AMD, the Day 0 support strategy reflects confidence that open-source models like Qwen will continue improving rapidly. By committing to immediate hardware support, AMD is signaling to developers that they can build applications on the latest models without worrying about whether their hardware will support them. This removes friction from the adoption cycle.

What Does This Mean for the Average User?

These developments are unlikely to affect casual users immediately. Running Qwen 3.8 27B requires either a high-end PC with an AMD Ryzen AI Max+ processor or a professional-grade graphics card. However, the trajectory is clear: as models become more efficient and hardware becomes cheaper, the barrier to entry will drop. Within two to three years, running advanced AI models locally could become as routine as running a web browser.

For developers and enterprises, the implications are more immediate. AMD's Day 0 support means they can start building applications on the latest open-source models without waiting for optimization cycles. Google's HEIR compiler means they can process sensitive data in the cloud while maintaining cryptographic privacy guarantees. Together, these tools expand the design space for AI applications, making it possible to choose between local inference, encrypted cloud inference, or hybrid approaches based on the specific requirements of each use case.

" }