DeepSeek R1 is one of the most capable open-source reasoning models available. This guide covers running the 7B, 14B, and 70B distilled variants on consumer hardware using Ollama or LM Studio.
DeepSeek R1 comes in several distilled variants based on Qwen and Llama architectures. The 7B distill runs on 8 GB VRAM at Q4 quantisation (5 GB model size), making it accessible on most gaming GPUs and MacBook Pros with 8 GB unified memory. The 14B distill delivers meaningfully better reasoning and needs around 10 GB VRAM.
For serious reasoning workloads — maths, coding, multi-step logic — the 70B distill is the sweet spot if you have 40+ GB of unified memory (M2 Max/Ultra, M3 Max) or a workstation with dual GPUs. The full R1 671B MoE model requires server-grade hardware and is best accessed via API.
DeepSeek R1 7B: 8 GB RAM/VRAM minimum, 12 GB recommended. Runs well on an M1/M2 MacBook Pro 8 GB or an RTX 3060 12 GB.
DeepSeek R1 14B: 12 GB VRAM minimum, 16 GB recommended. Works on an M2 Pro 16 GB or an RTX 3090/4070 Ti.
DeepSeek R1 70B: 40 GB unified memory (M2/M3 Max) or dual RTX 3090 (48 GB VRAM total). At Q4_K_M quantisation the model file is around 40 GB.
Install Ollama first if you haven't already (see our Ollama installation guide). Then in a terminal run: `ollama pull deepseek-r1:7b` for the 7B variant, or `ollama pull deepseek-r1:14b` for the 14B.
Start an interactive session: `ollama run deepseek-r1:14b`. DeepSeek R1 models emit chain-of-thought `<think>` blocks before their final answer — this is normal. The thinking section can be several paragraphs long for complex problems. The actual answer follows after the closing `</think>` tag.
To suppress the thinking output in the Ollama CLI, add `--nostream` and pipe through a quick filter, or use the API endpoint and strip the think tags programmatically.
Open LM Studio, search for 'DeepSeek R1' in the model browser, and select the appropriate size. The recommended quantisation is Q4_K_M for the best quality-to-size ratio. Download and load the model.
In the system prompt field, you can optionally add: 'You are a helpful AI assistant. Think through problems step by step before giving your final answer.' This primes the model to use its reasoning capability fully.
DeepSeek R1 excels at tasks with clear correct/incorrect answers: maths, logic puzzles, coding, and structured data extraction. Frame your prompts precisely — ambiguous prompts lead to unnecessary hedging in the thinking section.
For coding tasks, always include the programming language and any relevant context (existing functions, error messages, or expected behaviour). The model tends to write more idiomatic code when given clear examples of your coding style.
If responses feel slow, reduce the context length in the model settings. The default 32K–64K context loads more KV cache memory than most tasks need — for simple questions, 4K context is enough and will be 2–3× faster on CPU offload.
Ollama is the fastest way to run open-source LLMs locally. This guide walks you through installation on macOS, Linux, and Windows, pulling your first model, and running it from the terminal or via its REST API.
Read guideLM Studio is a desktop app for running local AI models without any terminal commands. This guide covers installation, downloading a model from Hugging Face, and starting a local inference server.
Read guideLlama 3.3 70B is Meta's best dense open-source model, matching GPT-4o on many benchmarks. This guide covers hardware requirements, quantisation selection, and optimal settings for running it on a high-end workstation or Apple Silicon Mac.
Read guide