The reasoning model that matched OpenAI’s o1 on math and code under an MIT license — and a genuine 671B mixture-of-experts that only a handful of home rigs on Earth can actually run whole.
Last verified 2026-07 · Page created; Q4_K_M/Q5_K_M/Q8_0 sizes verified against bartowski/DeepSeek-R1-GGUF. FP16 omitted (no standard single-file FP16 GGUF is published for a model this size).
DeepSeek-R1 is a reasoning-first model: it thinks in an explicit chain before answering, which is what let a Chinese lab match OpenAI’s o1 on math, code, and logic benchmarks — and release the weights under a plain MIT license, with no restriction on commercial use or distillation. That last part matters more than the headline number: DeepSeek used R1 to distill six smaller dense models (Qwen- and Llama-based, 1.5B to 70B), and those distills are how almost everyone actually runs "DeepSeek-R1" at home.
The full model is a 671B-parameter mixture-of-experts, but only 37B parameters activate per token — that active count is what drives generation speed, while the full 671B is what you have to fit in memory regardless.
This is the one model in this registry that is honestly out of reach for a single consumer PC. Even the smallest practical quant, Q4_K_M, is 404GB — that needs multiple 80GB datacenter GPUs (a 5–6x H100/A100 pool) or a very large unified-memory Mac cluster to hold, let alone run at speed. Community "run it on a single machine" builds exist, but they lean on aggressive sub-4-bit dynamic quants (down to ~130–150GB) and heavy CPU/RAM offload with correspondingly slow generation, not a normal desktop workflow.
For local use on realistic hardware — one to two consumer or prosumer GPUs, 16–48GB of VRAM — the practical move is DeepSeek-R1-Distill-Qwen-32B, which carries most of R1’s reasoning behavior in a 32B dense model that fits the same VRAM tiers as any other 32B-class model on this site.
Q4_K_M (404GB) through Q8_0 (713GB) scale roughly linearly with bits per weight, same as any other model — the difference here is that every rung on that ladder is still a multi-GPU server-class build. If you are pricing this out seriously, compare the cost of that hardware against renting an H100/H200 node by the hour for the (likely occasional) sessions where you need the full model rather than the distill.
| Quant | File size | VRAM @ 8K ctx | VRAM @ 32K ctx | 8 GB | 16 GB | 24 GB | 32 GB | 48 GB |
|---|---|---|---|---|---|---|---|---|
| Q8_0 | 713.3 GB | 916.8 GB | 1206.1 GB | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Q5_K_M | 475.4 GB | 643.2 GB | 932.5 GB | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Q4_K_M | 404.4 GB | 561.6 GB | 850.9 GB | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
Fit verdicts use the 8K context preset. Estimated from published weight sizes plus standard KV-cache math — see methodology.
No catalog GPU in our reference list fits this model yet — check the full AI-scored catalog.
Three ways to size a machine for DeepSeek-R1, from the smallest GPU that fits to the no-compromise option with headroom for the next model up.
| GPU | Bandwidth | Est. tok/s (Q4_K_M) |
|---|---|---|
| RTX 5090 | 1792 GB/s | ~58.5 |
| RTX 4090 | 1008 GB/s | ~32.9 |
| RTX 6000 Ada | 960 GB/s | ~31.3 |
| RX 7900 XTX | 960 GB/s | ~31.3 |
| RTX 5080 | 960 GB/s | ~31.3 |
| RTX 5070 Ti | 896 GB/s | ~29.2 |
Estimated from memory bandwidth ÷ active weight size — a ceiling, not a measured figure. Real throughput depends on the inference engine, batch size, and context length.
Owning the hardware is a one-time cost that keeps paying off the more you actually use it — every run after the first is free, the data never leaves your machine, and the GPU still does everything else a PC does (gaming, editing, the rest of your workload) between AI sessions. Renting cloud GPU time trades that upfront cost for pay-as-you-go pricing with no idle hardware to maintain, which wins for occasional, bursty, or much-larger-than-consumer-VRAM jobs. For a model that fits comfortably in the VRAM tiers above and gets used regularly, owning the GPU tends to be the cheaper path over time — for anything larger than a single card can hold, renting is often the more practical starting point.
No. Even the smallest common quant is over 400GB — that is multiple datacenter-class GPUs, not a desktop build. Run DeepSeek-R1-Distill-Qwen-32B instead, which fits normal 16–48GB VRAM tiers and carries most of the reasoning behavior.
DeepSeek-R1 is the full 671B mixture-of-experts. The Distill models (Qwen- and Llama-based, 1.5B-70B) are smaller dense models fine-tuned on R1’s reasoning traces — they are not quantized versions of R1, they are separate, much smaller models that inherit a lot of its reasoning style.
For any realistic local setup you would be running one of the distills through standard llama.cpp/Ollama/vLLM. Full R1 needs MoE-aware serving (vLLM, SGLang) on a multi-GPU or multi-node setup — this is server infrastructure, not a single-box local-LLM workflow.