Can you run it?

Install & run

Pull from the canonical deepseek-ai org on Hugging Face and pin the revision; the distribution is safetensors, so the load itself is safe. The realistic entry point for most adopters is a distill or a quantized build - the full V3/R1 weights need a cluster.

# Full weights via Transformers (needs a multi-GPU box)
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1", revision="<pin-a-commit>")
model = AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-R1", revision="<pin-a-commit>", torch_dtype="auto", device_map="auto")

# A distill on a modest GPU
ollama run deepseek-r1:8b            # community distill via Ollama

vLLM and llama.cpp are the other common paths; use a GGUF distill under llama.cpp/Ollama for CPU/small-GPU hosts. Verify the checksum after download.

Hardware requirements

The full V3/R1 is a ~671B-total / ~37B-active MoE: serving the full weights requires a multi-GPU cluster (high aggregate VRAM even though only ~37B params are active per token, because all experts must be resident). This is the single biggest practical barrier. The R1-Distill-Qwen/Llama checkpoints are dense models in the usual size range (e.g. 7B, 8B, 14B, 32B, 70B) and run on modest single GPUs - or, quantized, on a laptop. Pick the distill that fits your hardware; they are separate artifacts (see provenance).

Serving stacks

DeepSeek has first-class ecosystem support: vLLM and SGLang for GPU serving at scale, llama.cpp and Ollama for CPU/small-GPU and local use, and plain Transformers for scripting. R1's reasoning tokens are emitted inline (<think>), so whichever stack you choose, configure downstream parsing/stripping of the reasoning span before returning text to users.

Safe-deployment controls & Deployment Ceiling

Applying the playbook to DeepSeek:

• Supply chain: pull the canonical safetensors and pin the revision; the data-only format makes the load itself safe (no unsandboxed pickle). Prefer the canonical org over random distill uploaders.
• Legal: confirm MIT applies to your exact checkpoint. Recognise the systemic-risk Article 53/55 gap - as an EU deployer you cannot close it from upstream documentation.
• Safety: because no guard model ships, stand up your own input/output classifiers, prompt-injection defences, and a content/neutrality review step given the censorship behaviour, before any customer contact.
• Isolation: self-host - do not route sensitive data through DeepSeek's hosted API.

Deployment Ceiling: T2 (customer-facing, human-reviewed), conditional. The self-hosted open weights are supply-chain-safe (safetensors), and T2 is reachable with strong input/output guards, prompt-injection defences, and a neutrality/content review layer. T3+ requires heavy guardrails given the lighter safety tuning, and autonomy should be granted cautiously. T4 / EU high-stakes is not recommended, driven by two hard constraints: the Article 53/55 compliance gap for a systemic-risk model, and the censorship behaviour, which is disqualifying where topic-neutral factuality is required.

Available quantizations

An extensive community ecosystem of GGUF (llama.cpp/Ollama) and AWQ/GPTQ builds exists, alongside the R1-Distill family that is itself the most common way to shrink the footprint. Treat every quant/distill as a separate-trust artifact - its trust equals the uploader, not DeepSeek. Prefer builds from the canonical org or a host you trust, and verify checksums.

Fine-tuning & adaptation

The MIT license explicitly permits distillation and derivative training - this is why the R1-Distill family exists at all. Standard supervised fine-tuning and LoRA/QLoRA adaptation apply to the open weights (most practically to the smaller distills). If you distill or fine-tune, you own the resulting artifact's provenance and, as a downstream provider, may inherit your own EU AI Act obligations.

API / OpenAI-compatible integration

For application integration, self-host behind an OpenAI-compatible endpoint via vLLM or SGLang (both expose /v1/chat/completions-style APIs), so existing OpenAI-SDK code points at your own inference server. This keeps prompts on your infrastructure and is the recommended alternative to calling DeepSeek's hosted API when sensitive data is involved.