Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsIf a local language model runs slower than a review or benchmark suggests, first check where it is actually running and which phase is slow. A GPU in the computer does not prove the model is using it, and prompt processing and token generation can have very different speeds. Performance depends on the exact model and quantization, runtime and backend, hardware, context length, and workload—so there is no meaningful universal tokens-per-second expectation.
Why local LLM speed can differ from advertised results
A speed figure only describes the conditions under which it was measured. Differences in model file or quantization, runtime version, hardware, GPU placement, prompt length, context, output length, and concurrent requests can all make a comparison misleading. A short prompt’s processing rate is not comparable to generation speed after a long prompt, and a single request is not comparable to a concurrent serving workload.
| # | 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 |
As an Amazon Associate I earn from qualifying purchases.
There is no broadly representative cross-runtime figure that establishes a normal speed for local inference. Treat a review’s tokens-per-second number as a result for its particular setup, not a promise for yours.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchEstablish a baseline before changing settings
Record the conditions for one repeatable run. Keep the prompt and workload fixed, and note prompt-processing and generation timings separately if your runtime exposes them.
#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.
- Operating system, CPU, GPU or GPUs, system RAM, and GPU memory.
- Runtime and version or build, model name, and quantization.
- Context length, prompt length, output length, and number of concurrent requests.
- Whether other GPU workloads are active and whether the model was already loaded.
- Load time, actual CPU/GPU placement, prompt-evaluation rate, and generated-token rate.
Repeat the run under the same conditions before comparing it with a changed configuration. This makes it easier to tell a genuine improvement from ordinary variation or a different workload.
Confirm whether the GPU is actually doing the work
Do not infer GPU use from the presence of a graphics card. Check the runtime’s placement diagnostics and compare them with actual utilization and logs.
llama.cpp with CUDA
Check the startup log for GPU-offloaded layers and VRAM use. The llama.cpp performance troubleshooting guide says its example CUDA offload diagnostics indicate GPU use. If the expected offload is absent, verify that the binary was built with the required backend and that the GPU is available to the runtime.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Ollama
Run ollama ps and inspect the PROCESSOR column. Ollama’s FAQ documents output indicating GPU, CPU, or split CPU/GPU placement. Placement describes where model memory is loaded; if it does not match observed compute, also check GPU utilization and runtime logs.
Ollama in Docker
For containerized use, confirm that GPU access is passed through to the container. Ollama’s FAQ notes that NVIDIA GPU acceleration in Docker requires the relevant NVIDIA Container Toolkit setup.
Understand what a CPU/GPU split means
A split can indicate that the model does not fit entirely in available GPU memory, but it is not automatically a failure. The performance effect depends on the model, runtime, transfer behavior, and workload. Ollama says a model that fits on a single GPU typically minimizes transfers across the PCI bus; it does not say every split configuration is unusable.
To test whether memory capacity is the constraint, try a smaller model, a shorter context, or fewer concurrent requests, then compare placement and timings. Context and parallel requests add memory demand beyond the model weights. Ollama documents that parallel processing scales required memory with context length and request parallelism, and that models may be queued or unloaded when memory is insufficient.
For a clean baseline, close or unload other models and reduce concurrency. Keeping a model resident can avoid reload time for repeated requests, but that primarily changes startup or warm-response timing; it does not necessarily increase token-generation throughput.
Tune CPU threads without assuming more is better
Too many CPU threads can oversaturate the CPU and make generation slower. For llama.cpp, the troubleshooting guide recommends testing one thread when generation is very slow, then increasing gradually and comparing results. Find the useful setting for your own system rather than assuming one thread—or the largest possible count—is best.
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.
The guide’s example reports 5.5 tokens/s at one thread, 9.1 tokens/s at four, and 8.7 tokens/s at seven. Those figures came from an NVIDIA A6000 with 48 GB VRAM, a CPU with seven physical cores, 32 GB RAM, and a specified 30B 4-bit GGML model. The command and settings differ between rows, so this illustrates how threading and flags can interact; it is not a controlled thread-only comparison or a general benchmark.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Separate prompt processing from token generation
“Inference speed” can refer to more than one phase. Prompt processing happens before the model begins producing the answer; token generation follows. A long prompt can therefore delay the first token even when subsequent generation seems acceptable.
In its build documentation, llama.cpp says BLAS may improve prompt processing with batch sizes higher than 32, but does not affect generation performance. This is project-specific guidance, not a universal threshold for other runtimes. Benchmark the phase that matters to your use case instead of treating one rate as a complete measure of responsiveness.
Check the backend and build configuration
Runtime backends and build options affect which hardware is used and how work is scheduled. llama.cpp documents separate backends and performance options, but advanced flags are not interchangeable tuning recipes: some can increase memory use, reduce accuracy, or affect stability, while others trade speed for lower VRAM use in particular workloads.
Verify that the runtime build supports the backend you intend to use, then change one setting at a time. Avoid copying flags from a different project, GPU generation, or workload without checking what they do.
Change one variable at a time
Once placement and baseline timings are recorded, make a short log and test one change per run. Suitable variables include GPU offload, CPU thread count, context length, concurrency, model size or quantization, backend/build, and runtime version. Keep the workload fixed so you can identify which change affected prompt processing, generation, load time, or placement.
Recommended Free Tools
If testing a smaller model or context moves the model fully onto the GPU and improves the relevant timings, that is evidence of a memory-capacity constraint. Consider a hardware upgrade only after this diagnosis, and check the desired model’s memory needs and backend compatibility; advertised peak compute alone does not determine autoregressive generation speed.
How to compare local LLM benchmarks fairly
Compare like with like: use the same model file and quantization, prompt, context, output length, runtime version and backend, and concurrency. Record model residency and placement, load time, prompt-processing rate, and generated-token rate separately. Without those conditions, a tokens-per-second result is not a useful prediction of what your machine will deliver.
Ollama announced changes to model scheduling on September 23, 2025, describing more exact memory measurement and intended improvements to utilization and multi-GPU scheduling. That is a vendor announcement, not an independent speed study, so it does not establish a general percentage improvement or a guaranteed result for a particular setup.
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.
