Local AI can be slow on a legal document for reasons that have little to do with the language model itself. A scanned PDF may spend significant time in OCR; a long prompt takes more work and memory to process; the model may be loading from storage, running on the CPU instead of an accelerator, or competing with other requests for memory. Time each stage separately before changing settings or buying hardware.
Where the wait happens
A local document workflow usually has several stages: opening the file, extracting its text, running OCR if needed, preparing the prompt, loading the model, and generating the response. The first visible pause may happen before the model starts answering. Separate timings for these stages help identify whether you are dealing with document preparation, model inference, or both.
| # | 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.
- Record how long it takes to open the document and extract its text.
- If the PDF is scanned or image-based, time OCR separately.
- Note how long the model takes to load, how long it takes to produce its first token, and how long generation continues.
- Repeat the timings on a representative document and question. Keep the runtime, model, settings, and machine the same when comparing changes.
This is a diagnostic method, not a benchmark. Results depend on the computer, software, model, and document.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallIs the PDF scanned, or does it already contain text?
A PDF page may contain selectable text, an image of a page, or a mixture. Ordinary text extraction can read a usable text layer directly. An image-only page generally needs optical character recognition (OCR) before the model can work with its words.
#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.
PyMuPDF’s OCR guidance says OCR is roughly one thousand times slower than standard text extraction. That comparison explains why running OCR indiscriminately can add substantial preparation time; it does not mean every legal PDF will take a fixed amount of time to process. Check which pages actually need OCR, and retain the recognized text for reuse when the same document will be analyzed again. PyMuPDF recommends storing the OCR result in a TextPage rather than repeating OCR for the same page.
- For pages with a usable text layer, use ordinary extraction rather than OCR.
- For image-only pages, OCR only the pages that need it.
- Compare extracted text with the page image when exact wording, tables, stamps, handwriting, or layout matters.
OCR output may not preserve original font styling, and PyMuPDF’s guidance says Tesseract does not recognize vector graphics. Treat extracted text as a working representation, not a substitute for checking the page image when visual details affect interpretation.
Is the request too large for the task?
Sending an entire contract, case file, or statute can create a much longer request than asking about a relevant section. Long inputs use context capacity and can increase memory demand. In Ollama, the documented default context window is 2048 tokens; its documentation describes changing the context with /set parameter num_ctx or an API option. That default is specific to Ollama and may not describe another runtime or a customized setup.
Do not simply maximize context. A larger context can require more memory, and Ollama notes that parallel requests increase total context allocation. If available memory is insufficient, requests may queue.
For a question about a particular clause or issue, select or retrieve the passages that bear on it when your application supports that workflow. Preserve page numbers, section labels, and citations alongside the text so you can verify where an answer came from. There is no universally validated chunk size or overlap setting for dense legal text: the useful choice depends on the document, question, retrieval method, and model. Test your settings on representative documents rather than assuming one configuration fits every matter.
Is the model using the hardware you expect?
Model loading and generation depend on the model, runtime, available memory, and execution hardware. Microsoft’s Windows ML overview describes CPU, GPU, and NPU execution providers. It says discrete GPUs generally provide maximum performance for high-throughput generative-AI workloads, while NPUs are suited to battery-efficient sustained inference; Microsoft also cautions that results vary by hardware and model. This is platform guidance, not a benchmark for every runtime or legal-document task.
Check the runtime’s model-placement and memory information before concluding that a graphics card is being used effectively. Ollama documents ollama ps for inspecting loaded model placement. Also watch for queued requests and memory pressure: a model that does not fit comfortably, or several simultaneous requests with large contexts, can undermine responsiveness.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsRank #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.
On Windows, the available approaches are not interchangeable. Microsoft distinguishes built-in Windows AI APIs, ready-to-use local models through Foundry Local, and Windows ML for custom-model scenarios. Which option fits depends on the task, model, device support, and platform requirements.
For the llama.cpp SYCL backend specifically, its documentation warns that an Intel integrated GPU with fewer than 80 execution units will likely be too slow for practical use. This is a backend-specific warning, not a general threshold for other Intel GPU software, runtimes, or hardware.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Would a smaller or quantized model help?
Quantization can reduce the storage required for model weights, which may help when memory is a constraint. Microsoft’s Windows ML efficiency guide gives a basic comparison of four bytes per FP32 weight versus one byte per INT8 weight, while noting that actual savings depend on model structure and quantization method.
That storage comparison does not guarantee a particular increase in generation speed or a particular level of legal-domain accuracy. Compare the current model with a smaller or quantized candidate using the same representative documents and questions. Evaluate both latency and answer quality, including whether the model preserves distinctions, citations, and qualifications that matter to your work.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →When should you consider new hardware?
Consider hardware only after identifying the slow stage. A faster GPU will not remove time spent OCRing pages, and more context capacity will not correct poor extraction. Match a purchase to the bottleneck, the model you intend to run, available memory, runtime support, power and cost constraints.
The Council of Bars and Law Societies of Europe’s 2026 guide gives an illustrative local-inference setup of approximately €2,000, with 128 GB RAM and multiple lower-cost GPUs totaling 24 GB VRAM, describing it as capable of running 20–40B text-only models at a comfortable speed. These are guide examples based on September 2025 prices, not a current quotation or a universal performance guarantee; the guide warns that RAM prices are volatile. Its examples underline why memory capacity and bandwidth matter, but do not identify a universally suitable graphics card.
Quick Recap
A practical order of operations
- Measure the stages. Separate file opening and extraction, OCR, model loading, time to first token, and answer generation.
- Fix document preparation. Use the text layer where available, OCR only pages that need it, and reuse validated OCR output.
- Check placement and memory. Confirm which processor or accelerator is running the model, whether it fits, and whether other requests are competing for memory.
- Right-size the request. Provide the relevant passages and references when possible, and set context for the task rather than maximizing it by default.
- Compare model choices. Test smaller or quantized alternatives on the same legal questions, reviewing quality as well as speed.
- Then assess hardware. Choose only after identifying whether OCR, compute, memory, model fit, or platform compatibility is limiting the workflow.
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.
