STATION AI Max+ 395

Thunderobot's STATION AI Max+ 395 is a 2-liter mini workstation built around AMD's Ryzen AI Max+ 395, the 16-core "Strix Halo" APU that pairs Zen 5 CPU cores with a 40-CU RDNA 3.5 integrated GPU and a 256-bit LPDDR5X-8000 memory controller. The headline number is memory: 128GB of unified LPDDR5X, of which up to 96GB is addressable as VRAM by the Radeon 8060S. That single design decision is what makes this machine relevant for local AI work.

At $3,800 (26,999 CNY), this is prosumer/workstation territory, not consumer. It competes with other Strix Halo boxes, the NVIDIA DGX Spark at $3,999, and Apple's Mac Studio M4 Max 128GB. It does not compete with a dual-RTX-5090 rig on raw throughput, but it can hold models that a 32GB RTX 5090 physically cannot load.

For AI workloads the pitch is simple: this is one of the cheapest ways to get ~96GB of GPU-addressable memory in a 2-liter chassis that draws well under 200W at the wall.

Token generation for LLMs is memory-bandwidth-bound, not compute-bound. The LPDDR5X-8000 on a 256-bit bus delivers roughly 256 GB/s of theoretical bandwidth. That number determines your tokens per second, and it's the machine's primary limitation.

For context:

• RTX 5090: ~1,792 GB/s
• RTX 4090: ~1,008 GB/s
• Apple M4 Max: ~546 GB/s
• STATION AI Max+ 395: ~256 GB/s
• NVIDIA DGX Spark: ~273 GB/s

The STATION sits in the same band as the DGX Spark and roughly half of an M4 Max. That's fast enough to make 70B-class models usable, and slow compared to discrete GPUs on small models that already fit in 24–32GB.

The Radeon 8060S has 40 RDNA 3.5 compute units at up to 2.9 GHz, roughly 30 TFLOPS FP16. In practice it lands around RTX 4070 Laptop-class performance. Prefill (prompt processing) and image/video generation are compute-bound, so expect this GPU to feel like a mid-range laptop dGPU there, not a workstation card.

• Processor: AMD Ryzen AI Max+ 395, 16 Zen 5 cores / 32 threads, up to 5.1 GHz boost
• GPU: Radeon 8060S integrated, 40 CUs, 96GB shared memory
• Memory: 128GB LPDDR5X-8000, 256-bit bus, ~256 GB/s
• VRAM ceiling: 96GB addressable by the GPU
• Storage: 1TB PCIe Gen4 SSD, 2 free M.2 slots
• TDP: 132W (well above AMD's 55W default)
• Cooling: 4 heat pipes, 3 fans, aluminum mid-frame
• Network: 2x 10GbE RJ-45, Wi-Fi 7, Bluetooth 5.4
• Ports: 2x USB-C 40Gbps, HDMI 2.1, DisplayPort 1.4, SD UHS-II
• OS: Windows 11 Home (Chinese edition)

Sustained LLM inference should pull roughly 120–160W system power depending on load, about a tenth of a comparable dual-GPU workstation. The dual 10GbE ports are unusual at this price and matter if you're using the box as a small inference node on a lab network or pulling model weights from a NAS. Two free M.2 slots mean you can add fast local storage for model libraries without touching the boot drive.

This is where the machine earns its price. With up to 96GB of GPU-addressable memory, model size stops being the constraint.

• 70B-class dense at 4-bitLlama 3.1 70B, Qwen 2.5 72B, DeepSeek-R1-Distill-Llama-70B. Q4_K_M lands around 40GB, leaving ample room for KV cache.
• Mixtral 8x22B at Q4 (~80GB), loadable, tight with long context.
• Qwen 2.5 32B / Gemma 3 27B at 8-bithigh-quality quant with plenty of headroom.
• MultimodalQwen2.5-VL 32B, Gemma 3 27B (vision), Llama 3.2 Vision 11B all fit at 4–8 bit.
• MoE modelsQwen3 30B-A3B and similar run exceptionally well because only active parameters touch memory.

Estimates for llama.cpp with the Vulkan or ROCm backend, batch size 1:

• Llama 3.1 8B Q4_K_M: ~35–45 tok/s
• Qwen 2.5 32B Q4_K_M: ~10–13 tok/s
• Llama 3.1 70B Q4_K_M: ~4–6 tok/s
• Qwen3 30B-A3B (MoE, 3B active): ~40–60 tok/s

MoE is where this hardware shines, memory traffic scales with active parameters, not total parameters, so a 30B MoE with 3B active generates tokens like a 3B dense model.

Q4_K_M is the right default. It roughly halves memory traffic versus Q8 while keeping quality loss small for instruction-following and coding. On this hardware, Q8 doubles memory traffic and roughly halves token generation, only go there if you have VRAM to spare and need the quality.

Budget for KV cache on long-context work: a 70B model at 32K context with an FP16 KV cache adds several GB, and at 128K it can exceed 10GB. The 96GB VRAM budget absorbs this, but it isn't free.

Local chatbot and agent hosts. If you want a 70B-class model for coding assistance, research, or agentic workflows without paying per token or shipping prompts to an API, this is one of the few sub-$4K options that fits the model.

Developers building AI applications. Run the full stack locally: Ollama or llama.cpp for inference, a 70B model for heavy reasoning, smaller models for routing and tool calls, all on one box. Dual 10GbE makes it easy to expose as a service to the rest of a lab or home network.

Small teams running inference servers. A 2L chassis with 10GbE and 128GB unified memory is a plausible always-on node for a handful of concurrent users on a 32B model. It won't serve hundreds of simultaneous requests, but for a team of 5–20 it's viable.

Edge and space-constrained deployments. Two liters, low power draw, no discrete GPU. If you need LLM inference somewhere a rack server doesn't fit, this is a legitimate option.

Not for training. No NVLink, no CUDA, no HBM. LoRA fine-tuning a 7B is possible; anything larger is impractical. Treat this as an inference machine.

vs. NVIDIA DGX Spark (~$3,999). The closest direct competitor, 128GB unified memory, ~273 GB/s bandwidth, ARM CPU, CUDA. The Spark has the stronger software ecosystem (CUDA, TensorRT-LLM) and marginally more bandwidth. The STATION counters with a faster x86 CPU, dual 10GbE, and expandable M.2 storage. Pick the Spark if you need CUDA; pick the STATION if you want x86 binaries and 10GbE.

vs. Apple Mac Studio M4 Max 128GB (~$3,500–4,000). The Mac delivers ~546 GB/s, roughly 2x the bandwidth, so it generates tokens about twice as fast on the same model. But macOS limits you to MLX and llama.cpp, with no CUDA path. If you're already in the Apple ecosystem and only need inference, the Mac Studio is the better performer. If you need Windows, x86 tooling, or 10GbE, the STATION wins.

vs. a single RTX 5090 (32GB, ~$2,000+). The 5090 is 5–7x faster on models that fit in 32GB. It cannot run a 70B model at usable quality. Different tools for different jobs.

Availability: Pre-order in China now, sales starting October 9. Ships with Windows 11 Home (Chinese edition) and a 1-year warranty, budget time for an OS reinstall if you want English or a Linux/ROCm setup.