There is no universal hardware minimum for self-hosting an AI model. A small, quantized model may run on a CPU, while a larger model or a busy service may need a GPU—or multiple GPUs. Size the system for a specific model, context length, speed target, and number of simultaneous users, rather than relying on a single RAM or VRAM figure.
What determines the hardware you need?
Start with four choices: the model, its precision or quantization, the context length you need, and the workload. A model that fits in memory may still generate too slowly for your needs, and serving several requests at once can require more capacity than running a single response.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
MINISFORUM MS-02 Ultra Workstation Mini PC, Intel Core Ultra 9 285HX (24C/24T, up to 5.5GHz), PCIe... | $1,659.00 | Buy on Amazon |
| 2 |
|
GMKtec EVO-X2 AI Mini PC Ryzen Al Max+ 395 Superchip 128GB LPDDR5X 2TB SSD | $3,649.99 | Buy on Amazon |
- Model weights: Parameter count and numerical precision establish a rough starting point for memory.
- Context: Longer prompts and conversations require additional memory for the context or KV cache.
- Runtime: The inference software and backend allocate memory beyond the weights themselves.
- Performance and load: Acceptable latency, throughput, and concurrent requests influence the hardware choice.
NVIDIA’s local AI guidance recommends identifying target VRAM and performance requirements before choosing a model, and considering the operating system, model format, GPU architecture and memory, API needs, and throughput target when selecting a backend: NVIDIA: Build Local AI With NVIDIA GPUs.
How much memory do model weights need?
A useful rough estimate is parameter count × bytes per parameter. At BF16 or FP16, that is about two bytes per parameter for weights alone. It is a floor, not a complete system specification: runtime allocations and context memory add to it, and checkpoint formats affect the result.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
- High-Performance AI Processor：The MS-02 Ultra features an Intel Core Ultra 9 285HX (24C/24T, up to 5.5 GHz, 13 TOPS NPU), delivering fast and efficient performance for AI inference, algorithm development, and media workloads. A PCIe x16 expansion slot supports desktop-class GPU upgrades for advanced model training and accelerated computing tasks. It's ideal for creators, engineers, and teams handling intensive parallel workloads.
- 4 × M.2 PCIe 4.0 + 4 × DDR5 SODIMM slots:Four DDR5 SODIMM slots support up to 256 GB of memory, while ECC helps maintain data integrity in mission-critical environments. Four PCIe 4.0 M.2 slots support up to 24 TB of storage, supporting RAID 0/1/5/10, combining high-speed performance with data protection. It allows for the creation of independent scratch disks, media libraries, and project drives, providing high-throughput for production workflows.
- PCIe & USB 4.0 v2: Up to three PCIe slots can be equipped, including a dual-slot x16 GPU. The main slot supports PCIe 5.0, meeting the needs of high-bandwidth creative and computing workloads. USB 4.0 v2 (80Gbps) supports high-bandwidth external storage and displays.
- Ultra-fast Networking: Wi-Fi 7 further enhances wireless performance with next-generation speeds and low-latency stability. Intelligent bandwidth switching optimizes throughput in different network environments, ensuring optimal performance for enterprise or local networks. Dual 25GbE ports (providing up to approximately 3.125 GB/s bandwidth, about 25 times faster than traditional 1GbE), enabling seamless large-scale file transfers and parallel computing. 10GbE and 2.5GbE ports, with support for Intel vPro technology, ensure enterprise-grade remote management and deployment flexibility.
- Server-grade thermal architecture: Utilizing a dedicated CPU/GPU airflow design, equipped with a 6-pipe dual-fan cooler, it maintains stable performance even under sustained loads, delivering up to 140W Turbo power while maintaining a 100W TDP, and operating with noise levels as low as 36 dB. An integrated 350W power supply ensures stable and reliable output for demanding computing tasks and fully loaded extended configurations.
For a concrete but bounded example, Puget Systems measured just over 15 GB of VRAM for the BF16 weights of Meta Llama 3.1 8B Instruct. That is one model and test configuration, not a universal requirement for all 8-billion-parameter models: Puget Systems’ local LLM hardware primer.
“Model size” may refer to parameter count, the checkpoint’s disk size, or memory use while generating. These figures are not interchangeable. A checkpoint that occupies a particular amount of disk space does not guarantee that the same amount of VRAM will be sufficient.
How quantization changes memory needs
Quantization stores weights using lower-precision representations to reduce memory use. The llama.cpp documentation describes integer quantization options from 1.5-bit through 8-bit; the appropriate choice depends on the model, format, runtime, and acceptable quality trade-offs. Lower memory use does not, by itself, establish a particular speed or quality result.
In Puget Systems’ Llama 3.1 8B test, 8-bit and 4-bit versions used less VRAM than the BF16 version. The exact savings depend on the checkpoint and inference setup, so use the chosen model’s actual files and software guidance rather than assuming every quantized model has the same memory footprint.
Why context length and runtime add to the estimate
Memory use changes with context length; it is not just a weight-storage problem. Puget Systems reported that, in its test, context quantization and Flash Attention together brought VRAM use to 9.2 GB, compared with 28.6 GB when both optimizations were disabled. Those are results for that test configuration, not general sizing targets. The same testing found that Flash Attention reduced the memory impact as context grew: Puget Systems’ test discussion.
When estimating a build, include space for weights, context/KV cache, and runtime allocations. Keep ordinary system memory available for the operating system and other applications. If inference uses the CPU or offloads part of a model to it, model data also consumes system RAM and competes for CPU resources. The cited documentation does not establish one RAM multiplier that applies to every model.
Rank #2
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Can you run a local LLM without a GPU?
Yes. A discrete GPU is not required for every local inference setup. vLLM documents basic inference and serving on supported x86 and Arm CPU platforms, and llama.cpp supports CPU-plus-GPU hybrid inference that can partially accelerate models larger than available VRAM. Neither capability guarantees a particular response speed; whether CPU-only or hybrid inference is practical depends on the model and your latency expectations.
For CPU platform details, consult the vLLM CPU installation documentation. For supported backends and hybrid use, see the llama.cpp project documentation.
Free tools Windows power users keep installed
One-click scans. No signup required.
Which hardware path fits your use?
| Path | May suit | Main constraint |
|---|---|---|
| CPU-only | Small or quantized models, experimentation, and use where slower output is acceptable | System memory and CPU performance; documented support does not promise a universal speed target. |
| One GPU | Inference where the selected model and workload benefit from GPU acceleration | VRAM must accommodate weights, context, and runtime while meeting the desired performance. |
| CPU-plus-GPU hybrid or multiple GPUs | Models or workloads that exceed one GPU’s capacity | More involved allocation and performance trade-offs. |
| Apple Silicon unified-memory system | Local inference using a compatible backend that supports Apple hardware | Total shared memory and backend compatibility; llama.cpp lists Apple Silicon and Metal support. |
These are implementation paths, not performance rankings. Choose based on the model and workload, then check software support for your operating system, hardware, and model format. llama.cpp documents CPU/GPU hybrid inference, Apple Silicon/Metal support, and multi-GPU usage in its project documentation.
How to estimate a build before buying hardware
- Choose the model and model size. Use the specific checkpoint you intend to run; parameter count alone does not specify its final runtime footprint.
- Pick a precision or quantization. Check which formats your inference software supports and obtain the actual memory information for that model variant.
- Set the context length. Include the prompts and conversation history your application needs, not just the model weights.
- Define the workload. Decide how quickly responses should arrive and how many requests may run concurrently.
- Estimate total memory and compare compatible systems. Account for weights, context, runtime, and other system work; then compare usable memory, measured performance for the chosen workload, software support, power, noise, and budget.
- Validate with the intended application. Check current model and backend guidance, then measure memory use and speed in the setup you will actually use.
A 24 GB GPU is a hardware category, not a universal minimum or a promise that every model will fit. A capacity figure alone also cannot tell you whether inference will be fast enough. No specific GPU, price, or retailer listing is established here as the right choice for all self-hosters.
RAM and VRAM are not the same
VRAM is memory on a discrete graphics card; system RAM serves the CPU and operating system. GPU inference needs enough usable VRAM for the model and its runtime workload, while CPU inference and CPU offload draw on system RAM. Apple Silicon systems use unified memory, so capacity is shared across CPU and GPU work. The relevant question is not simply “How much RAM?” but which memory pool the chosen backend uses and how much the entire workload needs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Recommended Free Tools
