Specialist · sources checked 30 October 2026
DeepSeek R1 Distill Qwen 32B VRAM and RAM requirements
Run this model with Ollama or LM Studio ↓
Check a model download using its actual file size.
New to local AI? Find your GPU memory and understand these settings.
This checkpoint is not DeepSeek R1. The full R1 model in DeepSeek's release is a 671 billion parameter mixture-of-experts, and Ollama's deepseek-r1:671b tag is a 404 GB download with a 160K context. What this page estimates is DeepSeek-R1-Distill-Qwen-32B: a dense model distilled onto Qwen2.5 32B, published with the R1 paper in January 2025. Pulling deepseek-r1 without a size will not load it. On 30 October 2026 Ollama's default deepseek-r1:8b tag was DeepSeek-R1-0528-Qwen3-8B, a later and smaller distill, at 5.2 GB. The file that matches this page is deepseek-r1:32b, which is 20 GB and shares its digest with deepseek-r1:32b-qwen-distill-q4_K_M.
Distillation, in the sense DeepSeek means
DeepSeek generated reasoning traces with R1 and fine-tuned smaller dense models on about 800,000 of those samples. The 32B student started as Qwen2.5-32B. DeepSeek says the configs and tokenisers were changed slightly and asks you to use their settings. The repository and the distill weights are MIT. The base model was Apache 2.0, which DeepSeek records so that the lineage stays visible. MIT on the distill does not rewrite history for the Qwen weights it was trained from, but it is a far more ordinary commercial licence than the Llama community licence that covers DeepSeek's separate Llama 8B and Llama 70B distills.
The published config, from the Hugging Face commit that added it, is a Qwen2 causal language model: 64 layers, hidden size 5,120, intermediate size 27,648, 40 query heads, 8 KV heads, vocabulary 152,064, RoPE theta 1,000,000, sliding window off, and max_position_embeddings of 131,072. Head dimension is 128 because 5,120 divided by 40 is 128. That is the architecture behind the cache column. Qwen2.5 32B's own original window was shorter; the 131,072 figure is the limit DeepSeek wrote into this config. Their evaluation script does not use all of it. The paper caps generation at 32,768 tokens.
How the paper's scores were produced
DeepSeek's distill table is pass@1 estimated from 64 samples per question, at temperature 0.6 and top-p 0.95, with that 32,768 token generation cap. A single greedy answer will not reproduce the table. For the 32B Qwen distill the paper reports AIME 2024 pass@1 of 72.6 and cons@64 of 83.3, MATH-500 of 94.3, GPQA Diamond of 62.1, LiveCodeBench of 57.2 and a Codeforces rating of 1,691. On that same table the distill is ahead of the paper's o1-mini row on AIME 2024 (63.6), MATH-500 (90.0), GPQA Diamond (60.0) and LiveCodeBench (53.8), and behind it on Codeforces rating (1,820). "Beats o1-mini" is true for four of those five cells and false for the rating.
The smaller Qwen distills in the same table show what the extra parameters buy. The 14B scores AIME 2024 at 69.7, MATH-500 at 93.9, GPQA Diamond at 59.1, LiveCodeBench at 53.1 and Codeforces at 1,481. The 7B scores 55.5, 92.8, 49.1, 37.6 and 1,189. Moving from 14B to 32B is three points of AIME and a clearer gain on Codeforces, in exchange for roughly twice the memory. The 14B distill and the 7B distill are the honest alternatives when the 32B target below does not fit. Ollama lists the 14B Q4 package at 9.0 GB and the 7B Q4 package at 4.7 GB.
A 20 GB file is not a 24 GB success
Ollama lists the Q8_0 distill at 35 GB and the FP16 distill at 66 GB, beside the 20 GB Q4 tag. Our allowances, from the catalogue's 32.5 billion parameters, are 19.70 GB, 34.53 GB and 65.00 GB. The gap to Ollama's rounded packages is small. The gap that catches people is between the download and the running target, because a reasoning model spends its context on the trace.
| Precision | GPU VRAM target | System RAM target | Weight allowance |
|---|---|---|---|
| Q4_K_M | 24.6 GiB | 32 GiB | 19.70 GB |
| Q8_0 | 41.5 GiB | 48 GiB | 34.53 GB |
| FP16 | 76.2 GiB | 72 GiB | 65.00 GB |
At 8,192 tokens, Q4_K_M is already a 24.6 GiB GPU target, with 2.0 GiB of cache, 32 GiB of system RAM and a 40 GiB unified target. A nominal 24 GB graphics card is about 22.4 GiB once decimal gigabytes are converted, so this planning target sits above the card before a desktop compositor has taken anything. Even at 1,024 tokens the Q4 target is 22.7 GiB. "The 20 GB file fits in 24 GB" is a statement about the download. It is not a statement this calculator will support for a full-GPU load.
The paper's own 32,768 token generation cap is the more relevant stress test. At that length the Q4 cache is 8.0 GiB and the GPU target is 31.3 GiB, with a unified target of 48 GiB. A long trace is the normal output of this model, not an edge case. Budget for it or the load will fail halfway through a maths question that started small. Q8 at 8,192 tokens is a 41.5 GiB GPU target. FP16 is 76.2 GiB. Neither belongs on a single consumer card.
The config allows 131,072 tokens of position. At that length the Q4 GPU target is 58.0 GiB and the cache alone is 32.0 GiB. DeepSeek's Hugging Face example serves the model with --max-model-len 32768. That is the practical cap to copy unless you have measured a longer one. The context sweep for the first quantisation:
| Context | KV cache | GPU target | Mac unified target |
|---|---|---|---|
| 1,024 tokens | 0.25 GiB | 22.7 GiB | 32 GiB |
| 4,096 tokens | 1.00 GiB | 23.5 GiB | 32 GiB |
| 8,192 tokens | 2.00 GiB | 24.6 GiB | 40 GiB |
| 32,768 tokens | 8.00 GiB | 31.3 GiB | 48 GiB |
| 131,072 tokens | 32.00 GiB | 58.0 GiB | 80 GiB |
Known limits of a reasoning distill
There is no official non-thinking switch of the kind Qwen3 later shipped. The model is trained to emit a chain of thought. A short question can still produce a long trace, and that trace consumes the context and the KV cache you just sized. DeepSeek's paper discusses this style of generation for R1. Plan for the length even when you intend to hide the trace in the interface. The student is also not fresher than its Qwen2.5 32B base plus the R1 traces used to tune it. It will not know the world after that training cut, and it will sound sure while it is wrong.
Qwen3 8B in thinking mode, scored in Qwen's later technical report rather than in DeepSeek's paper, is ahead of this 32B on several of Qwen's rows: AIME 2024 76.0 against the 72.6 both papers print for the distill, AIME 2025 67.3 against 49.6, and IFEval 85.0 against 72.5. Qwen's LiveCodeBench v5 number for the distill is 54.5, which is not the same cell as DeepSeek's LiveCodeBench 57.2. Do not average them. Use DeepSeek's paper when you are judging this file on DeepSeek's protocol, and use Qwen's table when you are deciding whether an 8B thinking model makes the 32B unnecessary. Often, on a 12 GB GPU, it does. Qwen 2.5 32B Instruct is the base family this distill was built on, without the reasoning traces, and it is the checkpoint to prefer when you want a direct answer rather than a 32,768 token thought.
Sources for the distill
- DeepSeek R1 README: distill scores, 64-sample protocol, 800k fine-tune, MIT licence and the Qwen2.5 lineage
- DeepSeek-R1 paper, arXiv 2501.12948
- Hugging Face model card, including the vLLM example capped at 32,768 tokens
- Config commit: 64 layers, hidden size 5,120, 40 query heads, 8 KV heads, 131,072 positions
- Ollama deepseek-r1 readme: 32b is the Qwen 32B distill, and the default 8b tag is the later Qwen3 distill
- Ollama tags read 30 October 2026: 32b at 20 GB, Q8 at 35 GB, FP16 at 66 GB, 671b at 404 GB
Run DeepSeek R1 Distill Qwen 32B locally
These steps use Q4_K_M weights. Compare the Q4_K_M row above with your available memory before downloading. The command does not choose Q8 or 16-bit weights, even if you explored those in the calculator.
View the exact Ollama download, file size and licence. Download tags checked 16 September 2026; publishers can update their files. File size is not the same as total running memory.
Option 1: Ollama
- Install and start Ollama from its official website. Check its supported GPUs.
- Open Terminal on Mac or Linux, or PowerShell on Windows. Run the command below. On first use, Ollama downloads the model before opening a chat.
ollama run deepseek-r1:32b-qwen-distill-q4_K_MInside the Ollama chat, enter this setting before your first prompt to use the same 8,192-token context as the table above:
/set parameter num_ctx 8192Then try a short prompt such as “Explain how a solar panel works in three sentences.” In a second terminal, run ollama ps to inspect whether the model is on the GPU or partly on the CPU. CPU offload changes performance and falls outside this calculator's full-GPU assumptions. See Ollama's context and GPU-loading instructions.
Option 2: LM Studio
- Install LM Studio and open its model discovery screen.
- Search for DeepSeek R1 Distill Qwen 32B. Match the family, size and variant exactly. Check the publisher and model card; community conversions are not necessarily published by the original model creator.
- Choose a supported GGUF download labelled Q4_K_M to compare with our balanced estimate. Check the actual file size before downloading.
- Load the model into a chat. Set the context length to 8,192 tokens and check the GPU offload settings. Try a short prompt before adding a long conversation or document.
LM Studio's official model download guide. Search results and available formats can change; we have verified the Ollama tags, not every community LM Studio conversion.
If the model will not load
Check your runtime version and GPU support, then close other memory-heavy apps. Try a shorter context or a smaller model and recalculate. Confirm that the downloaded format matches the estimate. Extra conversations, vision processing and runtime-specific allocations can increase memory use.
Find a smaller model for my computer · Check my memory and settings