Higher-resolution modes preserve more page detail but increase visual processing and GPU memory pressure. If the current GPU repeatedly OOMs or forces aggressive compromises in batch size, move to a larger-VRAM NVIDIA GPU rather than lowering document quality blindly.
For dense tables, formulas, technical documents and small text that require Base, Large or dynamic crop modes.Hardwareanforderungen für DeepSeek OCR: GPU, VRAM, RAM und Speicherplatz
Erfahren Sie mehr über die Hardwareanforderungen von DeepSeek OCR für NVIDIA-GPUs, VRAM, RAM, CUDA, vLLM, PDF-OCR und die Hardwareauswahl für ZimaOS.
DeepSeek OCR hardware requirements at a glance
DeepSeek OCR is a GPU-oriented vision-language OCR model. The upstream repository defines a CUDA/PyTorch inference stack and GPU-based examples, while actual VRAM use changes with resolution mode, batching, vLLM settings and model version.
- GPU
- The official DeepSeek-OCR reference path uses NVIDIA CUDA. Its Transformers example explicitly moves the model to CUDA in bfloat16, and the published environment uses CUDA 11.8 with PyTorch 2.6.0.
- VRAM
- DeepSeek does not publish one universal minimum VRAM figure for the upstream model. ZimaSpace's current DeepSeek OCR(Nvidia GPU) App guidance specifies an NVIDIA GPU with at least 8 GB VRAM. Third-party estimates range from roughly 8–10 GB for lighter/Base use to 15 GB or more for less constrained full-precision workflows.
- System RAM
- No universal official system-RAM minimum is published. A Hugging Face community report for Base mode estimated roughly 10 GB RAM use, while a Medium implementation guide recommends 16 GB or more. Treat these as deployment observations rather than official requirements.
- Storage
- Plan SSD space for the model checkpoint, Python/CUDA environment, uploaded PDFs/images, rendered page images and OCR output. Third-party sources cite roughly 6.7 GB for the checkpoint and around 20 GB free as a practical starting point, but DeepSeek publishes no universal disk minimum.
- Resolution and throughput
- The official model supports Tiny 512px, Small 640px, Base 1024px, Large 1280px and dynamic Gundam modes. Higher resolution, batching and PDF concurrency increase GPU memory and compute demand.
- Best Zima starting point
- The current ZimaOS package is specifically DeepSeek OCR(Nvidia GPU), so ZimaBoard 2, ZimaCube 2 Standard and Pro are not complete local OCR solutions as shipped without a compatible NVIDIA GPU. ZimaCube 2 Creator Pack is the natural preconfigured GPU path, subject to verifying the exact installed GPU VRAM and container access.
From official requirements to the right setup
DeepSeek OCR sizing starts with the exact model version and image mode, then checks NVIDIA VRAM, CUDA/container access, system RAM, PDF concurrency and persistent document storage.
-
Official requirements
Confirm whether you are deploying DeepSeek-OCR or the newer DeepSeek-OCR-2. The original upstream repository and the ZimaOS App should not inherit OCR-2 third-party memory figures unless the installed package is verified to use that newer model.
-
Confirm your needs
Verify the GPU path before sizing around throughput. The official DeepSeek-OCR environment uses CUDA 11.8, PyTorch 2.6.0 and FlashAttention, and the Transformers example explicitly runs the model on CUDA in bfloat16.
-
Leave room to grow
Choose Tiny, Small, Base, Large or dynamic resolution according to document detail. Higher-resolution pages, crop mode, batching and concurrent PDF pages can raise VRAM use beyond a single-image Base-mode test.
-
Run it on ZimaOS
Install the ZimaOS NVIDIA package, verify nvidia-smi and container GPU visibility, process representative pages, then measure VRAM, system RAM, page latency and batch throughput before committing to large PDF queues.
Check every playback client
- DeepSeek-OCR versus DeepSeek-OCR-2 model version
- Exact NVIDIA GPU model and available VRAM
- CUDA driver and container GPU visibility
- Tiny, Small, Base, Large or Gundam resolution mode
- Single image, PDF batch or concurrent service workload
- Expected page resolution and document complexity
- SSD capacity for model, temporary pages and OCR output
- Other ZimaOS AI, storage and container workloads
Official minimum requirements
The DeepSeek-OCR upstream repository publishes a tested software environment and GPU inference examples, but it does not define one universal minimum GPU, VRAM, system RAM or disk capacity.
Use the upstream CUDA/PyTorch/Python stack as the official compatibility baseline. For the current ZimaOS DeepSeek OCR(Nvidia GPU) package, use the ZimaSpace-specific ≥8 GB VRAM requirement as the local deployment floor, then leave additional headroom for resolution, batching and other services.
| Requirement | Official minimum | What this supports |
|---|---|---|
| Official tested environment | CUDA 11.8 + PyTorch 2.6.0 | The original DeepSeek-OCR README explicitly states this tested environment. A newer driver/runtime may work, but should be validated against the selected vLLM or Transformers stack. |
| Python environment | Python 3.12.9 in the published conda setup | The upstream install example creates a Python 3.12.9 environment before installing PyTorch, vLLM requirements and FlashAttention. |
| GPU execution | CUDA-based NVIDIA path in the official Transformers example | The upstream example calls .cuda() and converts the model to bfloat16. This establishes the reference GPU path but does not publish a universal minimum NVIDIA SKU. |
| VRAM | No universal upstream minimum published | ZimaSpace separately states that its current DeepSeek OCR(Nvidia GPU) App needs an NVIDIA GPU with at least 8 GB VRAM. Third-party 8–15+ GB estimates should not be presented as DeepSeek's official minimum. |
| Supported image modes | Tiny 512, Small 640, Base 1024, Large 1280, plus dynamic Gundam | The official modes correspond to 64, 100, 256 and 400 vision tokens for the fixed-resolution modes. Larger/dynamic image processing can change memory and throughput. |
| Published throughput reference | About 2500 tokens/s for concurrent PDF inference on A100-40G | This is an upstream A100-40G vLLM reference, not a minimum requirement or a guaranteed page-per-second result for consumer GPUs. |
When to upgrade your hardware
Upgrade DeepSeek OCR hardware when VRAM, resolution or document concurrency becomes the actual bottleneck. OCR page count alone is not enough to size the GPU.
Base or Large mode approaches the GPU memory limit
PDF processing moves from one page at a time to batch service
You move from inference to fine-tuning or OCR-2 experimentation
vLLM can batch inputs and the upstream repository publishes a high-throughput A100-40G PDF example. Concurrency raises the amount of active image/model state, so a GPU that handles one page can still fail under production batch settings.
For document archives, multi-user OCR services and automated ingestion queues.The supplied 2026 OCR-2 guide suggests 8 GB VRAM for 4-bit inference, 16 GB for full precision and 24 GB or more for fine-tuning, but these are third-party OCR-2 figures rather than DeepSeek-OCR v1 official requirements.
For users moving beyond the current ZimaOS inference App into model experimentation, quantization or fine-tuning.Plan hardware growth with confidence
DeepSeek OCR scales cleanly when GPU memory, system memory, scratch storage and the long-term document archive are planned separately.
Keep the model and OCR scratch data on SSD
The model checkpoint is several gigabytes, and PDF workflows may render pages to images before inference and save markdown, bounding-box images or structured output afterward.
Use SSD or NVMe for the model environment, temporary page images and active OCR output; keep large source archives on separate capacity storage.Increase VRAM before chasing CPU cores
The reference implementation is CUDA-oriented and model execution is GPU-centric. More CPU cores help PDF preprocessing and I/O, but they do not solve CUDA out-of-memory failures.
Choose a larger compatible NVIDIA GPU when the measured bottleneck is VRAM or GPU throughput.Separate OCR compute from the document archive
OCR benefits from a fast GPU while long-term PDF/image retention benefits from scalable storage. These do not need to be the same resource tier.
Use ZimaCube 2 drive bays for source documents and results while keeping active model/runtime files on faster SSD storage.Control batch size and resolution before upgrading
Tiny, Small, Base, Large and dynamic modes trade visual fidelity against vision-token count and compute. Batch size, crop mode and vLLM memory settings can also change peak VRAM.
Measure the exact document pipeline first; reduce concurrency or choose a smaller mode only when accuracy remains acceptable.Can it run on ZimaOS?
ZimaOS currently lists a dedicated DeepSeek OCR(Nvidia GPU) App. ZimaSpace's own 2026 community release note describes it as requiring an NVIDIA GPU with at least 8 GB VRAM, so GPU availability is a hard deployment gate for this package.
Install the NVIDIA-specific DeepSeek OCR App
The current ZimaOS App Store lists DeepSeek OCR(Nvidia GPU) in the AI category. Treat it as a GPU-specific deployment rather than a generic CPU OCR container.
Open DeepSeek OCR(Nvidia GPU) in the ZimaOS App StoreVerify at least 8 GB NVIDIA VRAM and container visibility
ZimaSpace's current App announcement specifies NVIDIA GPU with at least 8 GB VRAM. Confirm the installed GPU, driver, nvidia-smi output and container device access before downloading the model or diagnosing application errors.
Read the ZimaOS AI App announcementDo not infer support from system RAM or Intel graphics alone
ZimaBoard 2 and non-Creator ZimaCube 2 configurations have capable CPUs and integrated graphics, but the current ZimaOS DeepSeek OCR package is explicitly the NVIDIA GPU edition. Add a supported NVIDIA GPU or use another deployment path instead of assuming Intel iGPU execution.
Read ZimaCube GPU expansion guidanceChoose Zima hardware for your DeepSeek OCR workload
For the current ZimaOS DeepSeek OCR(Nvidia GPU) App, the decisive requirement is a compatible NVIDIA GPU with sufficient VRAM. CPU, RAM and drive bays matter only after that GPU gate is satisfied.
Does the Zima system already have a compatible NVIDIA GPU with at least 8 GB VRAM?
No — the system is using only its built-in Intel graphics
Do not treat the current ZimaOS DeepSeek OCR(Nvidia GPU) App as ready to run. ZimaBoard 2, ZimaCube 2 Standard and Pro need a suitable NVIDIA GPU added through an appropriate expansion path or a separate GPU inference server.
- Compact host/controller only; local NVIDIA OCR requires added GPU hardwareZimaBoard 2 832
- More system RAM, but still requires an added NVIDIA GPU for this AppZimaBoard 2 1664
Yes — NVIDIA GPU path is present and verified
Use ZimaCube 2 when you also need integrated document storage and GPU expansion. Creator Pack is the preconfigured dedicated-GPU option; Standard and Pro require a separately selected compatible GPU before running the current NVIDIA App.
- Multi-drive OCR archive plus user-added compatible NVIDIA GPUZimaCube 2 Standard
- More CPU/RAM/network headroom plus user-added NVIDIA GPUZimaCube 2 Pro
- Preconfigured dedicated NVIDIA GPU path for local OCRZimaCube 2 Creator Pack
This is a workload guide, not a guaranteed pages-per-second or VRAM benchmark. Results depend on DeepSeek-OCR versus OCR-2, resolution mode, crop mode, precision, vLLM/Transformers settings, batch size, NVIDIA driver, container GPU access and other ZimaOS workloads.
| Zima hardware | Best for | Example workload | Core configuration | Recommended boundary | Next step |
|---|---|---|---|---|---|
| ZimaBoard 2 832 | A compact ZimaOS host for document storage, preprocessing or remote OCR workflows; not a complete current DeepSeek OCR(Nvidia GPU) system as shipped. | File ingestion, PDF preprocessing, automation and forwarding OCR jobs to a separate NVIDIA inference server. |
|
As shipped, it does not provide the NVIDIA ≥8 GB VRAM gate specified for the current ZimaOS DeepSeek OCR App. A custom GPU expansion setup or remote inference endpoint is required. | Get Now |
| ZimaBoard 2 1664 | A compact OCR workflow host with more system-memory headroom for preprocessing and other containers, but still requiring an external/added NVIDIA GPU for local DeepSeek OCR. | PDF conversion, automation, OCR queues and other ZimaOS services around a separate or expanded NVIDIA inference path. |
|
The 16 GB system-memory configuration may help preprocessing and service concurrency, but it is not itself a DeepSeek OCR GPU upgrade. | Get Now |
| ZimaCube 2 Standard | A multi-drive document archive and OCR host that will receive a separately selected compatible NVIDIA GPU. | Large PDF/image libraries, preprocessing, OCR output retention and local DeepSeek OCR after GPU expansion. |
|
Standard is not ready for the current DeepSeek OCR(Nvidia GPU) App as shipped. Eight GB system RAM also leaves less headroom for large PDF preprocessing and other services. | Get Now |
| ZimaCube 2 Pro | A stronger multi-drive OCR server with more CPU/RAM/network headroom and a user-added compatible NVIDIA GPU. | Batch PDF ingestion, OCR queues, document archive, other containers and high-speed file movement after GPU expansion. |
|
More CPU, RAM and 10GbE improve the surrounding document pipeline but do not satisfy the NVIDIA VRAM requirement by themselves. | Get Now |
| ZimaCube 2 Creator Pack | The most direct current Zima configuration for local DeepSeek OCR because it ships with a dedicated NVIDIA RTX PRO 2000 plus 64 GB system RAM and large SSD capacity. | Local document OCR, structured PDF-to-Markdown conversion, batch processing, AI services and large multi-drive source/result archives. |
|
The current ZimaCube 2 product page names RTX PRO 2000 but does not state VRAM in its text specification. Do not promise a fixed DeepSeek OCR batch size or throughput until the exact GPU memory and installed software stack are verified. | Get Now |
What the Press Says
Highlights from trusted reviewers worldwide.
“ZimaCube 2: Not just another NAS, tested with 25TB storage, local AI agents, 4K transcoding, and real homelab workflows.”Read full review
“The ZimaBoard 2 is a compact x86 server board that can be turned into a mini NAS, home server, media box, or self-hosting hub.”Read full review
“ZimaCube 2: A modern, high-performance NAS with plenty of room to grow—built for users who want more than basic storage.”Read full review
“Coverage focused on ZimaCube 2's open hardware foundation, no monthly fee, and self-hosting flexibility.”Read full review
Loved by the Community
Stories and reviews from people who build with Zima every day.
★★★★★
Zima Blade Little yet Powerful
Maybe I am not digital natives but I live with PCs since 12 years old in 1984 when IBM PC clone come to my home. Many years have passed and many operating system I've tried. For me Zima blade and CasaOS was a quantum leap for home PC enthusiast and server lab machine to make me stay curious and relevant for this era.
★★★★★
Very good!!
I use ZimaCube Pro as 5th Proxmox cluster node. It runs several VMs and containers, including a VM with GPU passthrough to run a self-hosted LLM. A specific LXC container runs a Samba server for NAS capabilities using four of six RAID 6 SATA HDDs with ZFS.
★★★★★
Great innovation for mini server!
It is very useful and makes a powerful mini server for many purposes, including university and college students in engineering and electronics. Thank you so much for making this server.
★★★★★
Avaliação ZimaBoard 2
Construí um servidor de uso pessoal. O desempenho está muito bom e funciona perfeitamente onde quer que eu esteja. A surpresa é não dependermos de grandes estruturas para termos nosso próprio servidor de dados. Como iniciante, estou gostando bastante do ZimaOS, pois ele é simples e eficiente.
Frequently asked questions
These answers separate DeepSeek's upstream requirements, ZimaOS package requirements and third-party VRAM estimates so deployment examples are not mistaken for one official minimum.
How much VRAM does DeepSeek OCR need?
DeepSeek's official repository does not publish one universal minimum VRAM number. For the current ZimaOS DeepSeek OCR(Nvidia GPU) App, ZimaSpace states that an NVIDIA GPU with at least 8 GB VRAM is required. Third-party estimates vary with precision and mode, so 8 GB should be treated as the Zima package floor rather than a guarantee for every workload.
Is 8 GB VRAM enough for DeepSeek OCR?
It can be enough for the current ZimaOS package floor and some lighter inference configurations, but headroom matters. The deepseek-ocr.io site says Base mode can run in roughly the 8–10 GB GPU class, while other implementations recommend around 15 GB or more for less constrained full-precision use. Batch size and image mode can push memory higher.
How much system RAM does DeepSeek OCR need?
No universal official system-RAM minimum is published. One Hugging Face user running Base mode on a 32 GB VRAM GPU estimated roughly 10 GB system RAM use, while the supplied Medium guide recommends 16 GB or more. These are observations, not formal DeepSeek requirements.
Can DeepSeek OCR run without an NVIDIA GPU?
The original official reference implementation is CUDA-oriented and the current ZimaOS App is explicitly labeled DeepSeek OCR(Nvidia GPU). Alternative community ports may exist, but they should not be used to claim that the current ZimaOS package supports CPU, Intel iGPU, AMD or Apple acceleration.
Can ZimaBoard 2 run DeepSeek OCR?
Not as a complete current ZimaOS DeepSeek OCR(Nvidia GPU) system using only the built-in Intel graphics. ZimaBoard 2 can host files, preprocessing and other workflow services, but local use of this ZimaOS App requires a compatible NVIDIA GPU with sufficient VRAM or a separate GPU inference server.
Which DeepSeek OCR mode needs the most hardware?
The official fixed-resolution modes progress from Tiny 512×512 to Small 640×640, Base 1024×1024 and Large 1280×1280, followed by dynamic Gundam tiling. Larger and dynamic modes process more visual detail and can increase GPU compute and memory pressure, especially with batching.
Is DeepSeek-OCR-2 the same hardware requirement as DeepSeek-OCR?
No. DeepSeek announced OCR-2 as a newer release in January 2026. The supplied DEV guide gives OCR-2 estimates of 8 GB VRAM for 4-bit, 16 GB for full precision and 24 GB or more for fine-tuning, but those are third-party OCR-2 figures and should not be automatically applied to the original ZimaOS DeepSeek OCR package.
When should I choose ZimaCube 2 Creator Pack for DeepSeek OCR?
Choose Creator Pack when you want the most direct preconfigured Zima path to a dedicated NVIDIA GPU plus 64 GB RAM, 1 TB system storage and multi-drive document storage. Verify the exact installed GPU VRAM and container access before promising a specific DeepSeek OCR resolution, batch size or throughput.
What sources and further reading informed this DeepSeek OCR hardware guide?
The official DeepSeek-OCR GitHub repository is the primary authority for the CUDA 11.8, PyTorch 2.6.0, Python 3.12.9, FlashAttention setup, supported resolution modes and A100-40G throughput example; it does not publish a universal minimum VRAM/RAM number. deepseek-ocr.io is an independent information site, not the canonical DeepSeek repository, so its 8–10 GB Base-mode and checkpoint-size claims are treated as third-party planning data. Hugging Face Discussion #31 is a community observation from one Base-mode deployment. The DEV article is specifically about the newer DeepSeek-OCR-2 and its 8/16/24 GB VRAM tiers are author guidance, not upstream DeepSeek-OCR v1 requirements. The Medium article's 15 GB+ VRAM, 16 GB RAM and 20 GB disk figures describe one implementation path rather than a universal minimum.
