A capable mini-PC for local AI inference in Australia costs between $600 and $1,200 and delivers a qualitatively different experience from running the same models on a NAS. Where a NAS generates 1 to 4 tokens per second on a 7B model, a mid-range mini-PC with a Core i5 or i7 processor generates 10 to 18 tokens per second. That difference changes local AI from a novelty you use occasionally to a tool you use daily. The right hardware tier depends on which model sizes you need to run, whether you want NPU-assisted inference, and whether you plan to run the device continuously or on demand.
In short: For most local AI use, a mini-PC in the $700 to $1,000 range with 32GB of RAM is the practical starting point. Entry-level N100 devices at $350 to $450 are too slow for regular interactive use. Core Ultra mini-PCs at $900 to $1,200 with NPU support are the best balance of capability and cost for running 7B to 13B models at conversational speed. Discrete GPU setups are only necessary for 30B or larger models.
What Determines AI Inference Speed on a Mini-PC
Token generation speed for LLM inference is determined primarily by two factors: the memory bandwidth available to load model weights from RAM into the processor, and the compute capability of the processor itself. For models that fit entirely in system RAM, memory bandwidth is often the bottleneck. Faster RAM (LPDDR5X vs LPDDR4X) and higher RAM capacity both contribute to inference speed. A model that does not fit in RAM and must page to disk becomes almost unusable.
Modern mini-PCs use three types of compute for inference. The CPU alone handles inference on older generation processors. The integrated GPU (iGPU) on Intel and AMD processors can accelerate inference using Vulkan or Metal on macOS. The NPU (Neural Processing Unit) on Intel Core Ultra and AMD Ryzen AI processors provides dedicated matrix multiplication hardware designed specifically for AI inference. Ollama and compatible runtimes can offload inference to the iGPU and NPU, which typically improves token generation speed by 30 to 80 percent compared to CPU-only inference on the same chip.
RAM capacity determines which model sizes are viable. A 7B parameter model at Q4_K_M quantisation requires approximately 4 to 5GB of RAM. A 13B model requires 8 to 10GB. A 32B model requires 18 to 22GB. A mini-PC with 16GB total RAM has approximately 12 to 13GB available for inference after the OS and background processes, which is sufficient for 7B models comfortably and 13B models at aggressive quantisation. 32GB of RAM is the practical minimum for running 13B models without compromise. See the guide on LLM quantisation levels for a full breakdown of RAM requirements by model size.
Mini-PC Hardware Tiers for Local AI
Australian mini-PC pricing for AI inference falls into four practical tiers. Entry-level devices use Intel N100 or N95 processors designed for efficiency, not compute throughput. Mid-range devices use older-generation Core i5 or i7 processors with stronger single-threaded performance. AI-capable devices use Intel Core Ultra or AMD Ryzen AI processors with dedicated NPUs. High-end devices add discrete GPU capability for running larger models at GPU-accelerated speeds.
| Entry tier (Intel N100/N95 class) | Approximate AU price: $350 to $450. Generates 3 to 6 tokens/sec on a 7B Q4 model. Suitable for occasional use and background tasks. Not recommended as primary AI hardware. |
|---|---|
| Mid-range (Core i5/i7 12th or 13th gen) | Approximate AU price: $550 to $750. Generates 10 to 18 tokens/sec on a 7B Q4 model. Handles 7B models comfortably at 16GB RAM; 13B viable at 32GB. Good starting point for regular AI use. |
| AI-capable (Core Ultra 5/7 or Ryzen AI) | Approximate AU price: $900 to $1,200. NPU and iGPU-assisted inference. Generates 15 to 30 tokens/sec on 7B models. Handles 13B models at 32GB RAM. Recommended for daily AI use. |
| High-end (Core Ultra 9 or discrete GPU) | Approximate AU price: $1,200 to $2,500. 30B to 70B models viable with GPU VRAM. Fastest local inference short of a full workstation GPU. Suitable for power users and multi-model workflows. |
Entry Tier: Intel N100 Class ($350 to $450)
Intel N100 and N95 mini-PCs are energy-efficient devices originally designed for light office use and media streaming. They are available in Australia from Amazon AU and some specialist retailers in the $350 to $450 range and are often marketed as capable of running local AI. Technically they can run Ollama and serve a 7B model. Practically, the inference speed is 3 to 6 tokens per second at Q4_K_M quantisation, which means a typical response to a prompt takes 30 to 60 seconds. That is tolerable for background summarisation tasks but frustrating for interactive conversation.
The N100 tier is worth considering only if you already own one for another purpose (as a home server, for Docker containers, or as a media box) and want to experiment with local AI as a secondary function. It is not worth buying an N100 device specifically for AI inference when the next tier up costs $200 more and delivers three times the speed.
If budget is the primary constraint and $450 is the ceiling, consider a second-hand 12th-generation Core i5 or i7 mini-PC rather than a new N100 device. Older generation Core processors significantly outperform the N100 for AI inference workloads even at equivalent price points in the used market.
Mid-Range: Core i5/i7 12th and 13th Gen ($550 to $750)
Mini-PCs using Intel Core i5 and i7 processors from the 12th and 13th generations represent the current value sweet spot for local AI. These processors have strong single-threaded performance and memory bandwidth that translates directly to faster token generation. At 16GB RAM, a 13th-generation Core i7 generates 10 to 15 tokens per second on a 7B Q4 model. At 32GB RAM, the same hardware handles 13B models at 6 to 10 tokens per second.
Key buying criteria for this tier: look for models with upgradeable RAM rather than soldered memory. Some mini-PCs at this price point solder the RAM to the motherboard, which limits you to the factory configuration. Upgradeable SO-DIMM slots allow expanding to 32GB or 64GB as needed. Also check for two M.2 NVMe slots where possible: one for the OS drive and one for model storage, which keeps model access fast without requiring external drives.
Australian availability for this tier is good. Brands including Beelink, Minisforum, and GMKtec sell through Amazon AU, with some specialist retailers also carrying stock. Local retailer pricing may be 10 to 20 percent higher than Amazon AU for the same devices, though Australian Consumer Law protections apply to purchases from Australian retailers and may be worth the premium for buyers who prefer local support.
AI-Capable: Core Ultra and Ryzen AI ($900 to $1,200)
Intel Core Ultra 5 and Core Ultra 7 processors (formerly Meteor Lake, now Lunar Lake) include a dedicated NPU alongside a stronger integrated GPU. The NPU is specifically designed for matrix multiplication workloads at low power draw. Ollama and compatible runtimes can offload inference tasks to the NPU and iGPU simultaneously, which improves token generation speed compared to CPU-only inference without the power cost of a discrete GPU.
In practice, a Core Ultra 7 mini-PC at 32GB LPDDR5X RAM generates 15 to 25 tokens per second on a 7B Q4 model and handles 13B models at 8 to 14 tokens per second. That puts conversational responses in the 3 to 8 second range for a typical 100-token reply, which is comfortable for daily interactive use. AMD Ryzen AI processors in this price range (the Ryzen AI 9 HX class) offer comparable performance with slightly different memory architecture characteristics.
This is the recommended tier for anyone buying a mini-PC specifically for regular local AI use. The Core Ultra and Ryzen AI generation represents a meaningful step above 12th and 13th generation Core processors for inference workloads, not just a minor spec bump. The $200 to $400 premium over mid-range devices is justified if you plan to use local AI daily or need 13B models to run at interactive speed.
High-End: Discrete GPU and Core Ultra 9 ($1,200 to $2,500+)
The high-end mini-PC tier includes devices with discrete GPU options, typically via an external GPU enclosure or mini-PCs with integrated discrete graphics. A dedicated GPU with 8GB or more of VRAM changes the inference equation significantly. When a model fits entirely in GPU VRAM, the GPU's high memory bandwidth handles inference far faster than system RAM. An RTX 4060 with 8GB VRAM generates 40 to 80 tokens per second on a 7B model. A 13B model at Q4 requires approximately 8GB VRAM, putting it just within that tier at the boundary.
Core Ultra 9 processors at the high end of the integrated-only range generate 25 to 35 tokens per second on 7B models using combined NPU, iGPU, and CPU inference. They also support larger RAM configurations (up to 96GB LPDDR5X in some models), which makes 70B models at aggressive quantisation theoretically viable in system RAM, though the inference speed at that scale is slow.
For most home users, this tier is more hardware than the use case justifies. The jump from a Core Ultra 7 at $1,000 to a discrete GPU setup at $2,000 doubles the cost and the power draw for an improvement in token speed that only matters if you are running models above 13B parameters regularly.
Mini-PC AI Performance by Tier (7B Q4_K_M Model, Approximate)
| N100 Entry | Core i7 13th Gen | Core Ultra 7 | Discrete GPU (8GB VRAM) | |
|---|---|---|---|---|
| Approximate AU price | $350 to $450 | $550 to $750 | $900 to $1,200 | $1,500 to $2,500+ |
| 7B model tokens/sec | 3 to 6 | 10 to 18 | 15 to 25 | 40 to 80 |
| 13B model tokens/sec | Not viable | 6 to 10 (at 32GB) | 8 to 14 (at 32GB) | 20 to 40 (in VRAM) |
| Max practical model size | 7B only | 13B at 32GB RAM | 13B comfortable, 30B possible | 30B to 70B with enough VRAM |
| NPU support | No | No | Yes | Varies by model |
| Idle power draw | 8 to 12W | 15 to 25W | 20 to 35W | 25 to 60W |
| Under inference load | 15 to 25W | 35 to 55W | 45 to 70W | 100 to 250W (GPU) |
| Recommended for | Experiments only | Daily 7B use, occasional 13B | Daily AI primary device | Multi-model, 30B+ models |
What to Look for When Buying
RAM capacity and upgradeability are the most critical factors. Confirm whether the device has upgradeable SO-DIMM slots or soldered memory before purchasing. 32GB is the practical minimum for 13B model use. 16GB is adequate for 7B models with headroom. Soldered 16GB with no upgrade path locks you to 7B models permanently.
Storage configuration matters for model management. LLM model files are large: a 7B model at Q4 is 4 to 5GB, and storing five or six models takes 25 to 35GB. A 256GB SSD is too small once the OS and model library are installed. 512GB is the practical minimum; 1TB is preferred if you plan to keep multiple models. Some mini-PCs include two M.2 slots, allowing a separate drive for model storage without replacing the OS drive.
Thermal performance under sustained load is important for AI inference, which runs the processor at high utilisation for long periods. Check whether the device has an active fan or is fanless. Fanless mini-PCs are quiet and suitable for desk use, but they throttle more aggressively under sustained load than fan-cooled models. For background inference on a NAS-adjacent server, a fan-cooled model that can sustain higher clock speeds without throttling is preferable to a fanless device that throttles to manage heat.
Where to buy in Australia: Amazon AU is the most accessible source for mini-PC brands not carried by specialist retailers, with regular stock and competitive pricing. Scorptec and PLE carry some mini-PC brands alongside NAS hardware. Australian Consumer Law applies to purchases from Australian retailers, which provides protections for warranty and returns that international purchases may not guarantee.
Related reading: our NAS buyer's guide, our NAS vs cloud storage comparison, and our NAS explainer.
Use our free AI Hardware Requirements Calculator to size the hardware you need to run AI locally.
What is the minimum RAM for running local AI on a mini-PC in Australia?
16GB is the minimum for running 7B parameter models at Q4_K_M quantisation with adequate headroom for the OS and background services. A 7B Q4 model requires approximately 4 to 5GB of RAM, leaving 10 to 11GB for other processes on a 16GB system. 32GB is strongly recommended if you want to run 13B models or keep multiple smaller models loaded simultaneously. See the RAM tier guide for a full breakdown of which models run at each memory level.
Should I buy a mini-PC or use my existing NAS for local AI?
If you already own a NAS with 8GB or more of RAM, start there. Ollama runs on Synology, QNAP, and most NAS hardware that supports Docker. NAS-grade AI inference runs at 1 to 4 tokens per second on a 7B model, which is usable for background tasks and occasional use. If you find yourself wanting faster interactive responses, the investment in a mid-range mini-PC ($600 to $800) delivers a 5 to 10 times speed improvement. Many users run both: NAS for storage and background AI tasks, mini-PC as the primary AI inference endpoint. See the full mini-PC vs NAS comparison for a detailed breakdown.
Does a mini-PC need a discrete GPU to run local AI?
No. Modern Core Ultra and Ryzen AI mini-PCs run 7B and 13B models at conversational speed using their integrated GPU and NPU without any discrete graphics card. A discrete GPU becomes valuable for 30B or larger models, where GPU VRAM capacity changes what is possible at practical inference speeds. For most home users running 7B or 13B models, a capable integrated graphics mini-PC in the $900 to $1,200 range is more than sufficient. A discrete GPU setup more than doubles the cost and significantly increases power draw for a speed improvement that matters only above 13B parameters.
Is it better to buy a new mini-PC or a second-hand workstation for local AI?
A second-hand desktop workstation with an older discrete GPU (RTX 3060 or RTX 4060 class) can outperform a new mid-range mini-PC for inference speed at comparable or lower cost, particularly for 13B and larger models where GPU VRAM matters. The trade-off is power draw and physical size. A desktop with an RTX 3060 draws 150 to 250 watts under inference load versus 40 to 60 watts for a mini-PC. For always-on setups, that difference adds $100 to $300 per year in electricity depending on your state's rates. For a device that runs only during active use, the power difference is less relevant, and the performance advantage of a GPU workstation may justify the size and noise.
Can a mini-PC run local AI as a server for the whole household?
Yes. Ollama on a mini-PC exposes an API on port 11434 that any device on the network can reach. Combined with Open WebUI running in Docker, any phone, tablet, or computer on the household network can access the local AI interface through a browser without installing anything. A Core Ultra mini-PC with 32GB RAM handles multiple simultaneous requests, though concurrent inference from multiple users simultaneously will queue requests and reduce per-user throughput. For a household of two to four people using AI at different times rather than simultaneously, a single capable mini-PC is sufficient. See the Open WebUI setup guide for configuration steps.
Comparing a mini-PC against running AI on your existing NAS? The full comparison covers performance, cost, and which platform suits which use case in Australia.
Read the Mini-PC vs NAS Comparison