The mini PC market continues to evolve rapidly, with manufacturers like Beelink consistently pushing the boundaries of performance and features within compact form factors. The latest iteration, the Beelink Ser 10 Max, introduces AMD's cutting-edge "Gorgon Point" Ryzen AI 9 HX470 processor, promising significant advancements over its predecessors. This review delves into the Ser 10 Max, comparing it against last year's Ser 9 and earlier models, as well as the Apple Mac Mini M4 Pro, to assess its performance in various real-world and synthetic benchmarks, with a particular focus on its AI capabilities.

Hardware Overview and Key Differentiators

The Beelink Ser 10 Max is built around the AMD Ryzen AI 9 HX470 processor, codenamed "Gorgon Point." This chip represents a generational leap from the "Strix Point" found in the Ser 9, featuring a hybrid architecture combining Zen 5 and Zen 5c cores.

Beelink Ser 10 Max Specifications:

• CPU: AMD Ryzen AI 9 HX470 (Gorgon Point)
  • 12 Cores / 24 Threads
  • Zen 5 & Zen 5c Architecture
  • Boost Clock up to 5.2 GHz
• AI Acceleration:
  • MPU (Ryzen AI)
  • iGPU: Radeon 890M
  • NPU: XDNA 2 (55 TOPS)
  • Combined AI Performance: Up to 86 TOPS
• Memory: 32GB / 64GB / 96GB DDR5 5600 (User-configurable and upgradeable)
• Storage: User-upgradeable (supports up to 8TB SSDs)
• Connectivity: USB4, HDMI 2.1, DisplayPort 1.4, 10 Gigabit Ethernet
• Display Support: Triple 4K monitor support (up to 240Hz on one display)

Comparison Points:

• Beelink Ser 9: Features the previous generation AMD "Strix Point" chip. While still a capable machine, it lacks the latest CPU architecture and the dedicated NPU found in the Ser 10 Max. It typically offers 2.5 Gigabit Ethernet and has limitations on RAM upgradeability.
• Beelink Ser 8: An older generation model, serving as a baseline for generational improvements.
• Apple Mac Mini (M4 Pro/Base): A premium competitor. The base M4 model is limited to 24GB of RAM, which is insufficient for many modern AI and development workflows. The M4 Pro models offer more RAM but come at a significantly higher price point, especially when configured with higher storage and networking options.

The Ser 10 Max also offers a visually distinct orange variant pre-installed with OpenClaw, albeit at a $100 premium. Notably, configurations like 96GB RAM with 2TB SSD are priced competitively against similarly specced Mac Minis, which often cost more and offer less RAM.

Performance Benchmarks: Web, Development, and General Computing

To gauge the Ser 10 Max's performance, a series of benchmarks were conducted, comparing it against its predecessors and the Mac Mini M4 Pro.

Web and JavaScript Performance

• Speedometer 3: This benchmark simulates real-world web application interactions. The Ser 10 Max achieved a score that significantly closed the gap with Apple's M4 Pro, demonstrating strong single-threaded performance.
• Web Tooling Benchmark (V8): Testing JavaScript toolchains like TypeScript, Babel, and Terser, the Ser 10 Max showed substantial improvements over the Ser 9. Its TypeScript score saw a 65% jump, and the geometric mean increased by over 50%. Crucially, the Ser 10 Max's performance is now within 5% of the Mac Mini M4 Pro on TypeScript, and only trails by 11-13% on the geometric mean, a remarkable improvement from the Ser 9's 75% deficit.

General CPU Performance

• Geekbench 6: While often criticized as synthetic, Geekbench provides a widely recognized metric. The Ser 10 Max scored 2993 in single-core and 15216 in multi-core. This represents a significant leap over the Ser 9 and nearly matches the M4 Pro Mac Mini in multi-core performance (less than 1% difference), indicating strong parallel processing capabilities.

Developer Workloads

• Python Build Test: This test pushes all CPU cores to their maximum with an interpreted language. The Ser 10 Max completed the test in 28.9 seconds, virtually identical to the Ser 9 (28.64 seconds) and still faster than the base M4 Mac Mini (31.41 seconds). While not a generational leap here, it maintains its lead over the base Apple offering.
• NX Mono Repo Build: This test, simulating JavaScript development at scale with process spawning and file system operations, showed an astonishing improvement. The Ser 10 Max completed the build in 2.91 seconds, a five-fold increase over the Ser 9 (14.1 seconds). While the test's drift means direct comparison to older results is difficult, the sub-3-second build time is exceptionally fast for a mini PC.
• .NET Build (Synthetic): A custom .NET build designed to stress the compiler. The Ser 10 Max scored 90.9 seconds, a negligible 0.1-second difference from the Ser 9 (91 seconds). This result suggests that while the CPU architecture is new, this specific synthetic workload did not reveal a significant generational advantage for compiled code. However, it still outperforms the base M4 Mac Mini (106.7 seconds).
• Umbraco CMS Build (Real-world .NET): Compiling a mature, open-source .NET project like Umbraco is I/O intensive. Here, the Ser 10 Max clocked in at 161 seconds, which was slightly slower than the Ser 9 (149 seconds) and matched the Ser 8. This outcome highlights how I/O bottlenecks can overshadow CPU improvements in real-world scenarios, and where Apple's silicon, particularly the M4 Pro (84 seconds), excels. It's also noted that Windows Defender can significantly impact compilation times, with disabling it improving Ser 10 Max results.

AI and LLM Performance: Leveraging Multiple Accelerators

The Beelink Ser 10 Max's most significant upgrade lies in its integrated AI hardware: the CPU, the Radeon 890M iGPU, and the XDNA 2 NPU. AMD claims a combined 86 TOPS of AI performance.

Local LLM Inference

Running local Large Language Models (LLMs) on Windows often defaults to the CPU, which is inefficient.

• Initial Test (CPU Only): Using Ollama with a Qwen 2.5 7B model, the CPU was heavily utilized, while the GPU and NPU remained idle. This resulted in slow inference speeds.
• GPU Acceleration (Vulkan): By enabling the Vulkan flag for Ollama, the workload shifted to the Radeon 890M iGPU. This yielded substantial speed improvements:
  • Llama 3 8B: Increased from 27 to 37.5 tokens/sec.
  • Qwen 2.5 1.5B: Increased from 47.5 to 68.3 tokens/sec.
  • The iGPU showed 100% utilization during these tests.
• NPU Acceleration (Hybrid Mode): The XDNA 2 NPU, designed for specific AI tasks, was tested using Lemonade Server. In a hybrid mode where the NPU handles the "prefill" stage and the iGPU handles "decode," the NPU demonstrated its strength.
  • For a 7B model, the NPU achieved 631 tokens/sec in hybrid mode, which is 2.5 times faster than the iGPU alone (240 tokens/sec) for the same task.
  • The NPU excels in tasks involving long contexts, such as Retrieval Augmented Generation (RAG), agents, or coding assistance, where its efficient prefill stage is beneficial. The iGPU remains superior for streaming chat-like interactions.

Unified Memory Architecture (UMA) and Large Models

A surprising capability demonstrated is the Ser 10 Max's ability to load LLM models larger than its dedicated iGPU VRAM.

• Loading Large Models: The Radeon 890M iGPU has a reported 4GB of dedicated memory. However, the Ser 10 Max successfully loaded an 8.37GB Q4 quantized 14 billion parameter model onto the iGPU. This is achieved through AMD's Unified Memory Architecture (UMA), where the iGPU can dynamically draw from system RAM when its dedicated VRAM is insufficient.
• Performance: This allowed for prompt processing at approximately 50 tokens/sec and generation at 8.8 tokens/sec for the 14B model, with over 20GB of iGPU memory in use (combining dedicated and shared). This UMA capability suggests that even larger models (potentially 22B or 30B parameters with Q4 quantization) could run on the iGPU by leveraging system RAM.
• Ser 9 Relevance: Importantly, the Ser 10 Max shares the same Radeon 890M iGPU as the Ser 9. This means that users who purchased the Ser 9 for AI tasks can still leverage its hardware for running larger models by updating their software and utilizing UMA effectively.

Recommendations and Conclusion

The Beelink Ser 10 Max represents a significant step forward, particularly in its AI processing capabilities and overall performance uplift over previous generations.

Buy the Ser 10 Max if:

• You require the dedicated NPU for AI workloads like RAG, agents, or long-context processing.
• User-upgradeable RAM is a priority, offering flexibility beyond the Ser 9's limitations.
• The inclusion of a 10 Gigabit Ethernet port is essential for your network setup.
• You are looking for a powerful mini PC that significantly closes the performance gap with premium competitors like Apple's Mac Mini in many development and AI tasks.

Consider skipping the Ser 10 Max if:

• Your primary workload is LLM inference using the iGPU, as the Ser 9 offers the same iGPU hardware. In this case, updating software to leverage UMA and NPU features might suffice.
• Budget is a primary concern, and the performance gains over the Ser 9 do not justify the cost for your specific use case. Waiting for the next generation ("Medusa" in 2027) might be a better option.

In conclusion, the Beelink Ser 10 Max is a compelling mini PC that successfully integrates AMD's latest "Gorgon Point" architecture. It offers substantial improvements in CPU and AI performance, making it a strong contender for developers, AI enthusiasts, and power users seeking a compact yet potent computing solution. Its competitive pricing, especially when compared to Apple's offerings, further solidifies its position in the market.