whatmodel.sh / DeepSeek R1 Distill Qwen 32B

DeepSeek R1 Distill Qwen 32B

+ contribute

DeepSeek R1 Distill Qwen 32B is an open-weights language model from DeepSeek, released January 2025. Its published size is 32.8B parameters and a 131k-token context window. Licence: mit.

55.9±0.9#246 of 719 LLMsfrom 28 benchmark results across 11 sources

$ how it compares

The LLMs ranked either side of DeepSeek R1 Distill Qwen 32B, out of 828 with a score, then more from DeepSeek.

#	model	score	where	RAM	released
295	Qwen3-VL 4B Thinking Alibaba	56±1	local	~4.1GB	2025-10
296	EXAONE Deep 32B LG AI Research	56±2	local	~22GB	2025-03
297	DeepSeek R1 Distill Qwen 32B DeepSeek	56±1	api $1.59/M +local	~21GB	2025-01
298	Qwen2.5-VL 72B Instruct Alibaba	56±1	local	~46GB	2025-01
299	Granite 4.1 30B IBM	56±1	local	~19GB	2026-04
	more from DeepSeek
21	DeepSeek V4.1 Flash DeepSeek	78±1	api $0.26/M +local	~537GB	2026-09
31	DeepSeek V4 Pro (0813) DeepSeek	76±1	api $1.63/M +local	~1010GB	2026-08
35	DeepSeek V3.2 Speciale DeepSeek	76±3	comparison only

see all LLMs · find the ones that fit your machine

$ what it's good at

coding▁▁▁▃▆██▇▆▄▃▁ #247 of 458 · 19% of its score · 8 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails QwQ-32B-Preview, InternLM2.5 20B Chat, Mistral Small Instruct 22B (2409) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench (version unspecified) lab	57.2% ✎	▄····▄▄··▄█▄ beats 4/6	medium · avg 41%	▇▇ 5.9%	at its level
WebApp1K-Duo-React pass@1 (onekq-ai)	8.7% ✎	▃··▁·▁▁▁▆█▇█ beats 0/45	hard · avg 32%	▇ 3.4%	at its level
WebApp1K-React pass@1 (onekq-ai)	32.3% ✎	▃▁▁█▁▄▅▃▄▇▁▄ beats 7/31	medium · avg 44%	▇ 2.8%	sharp · at its level · noisy
CodeReviewQA: change localisation (hard), invariant accuracy	64.6% ✎	█▂▁▁▂▂▂▃▅▅▅▇ beats 44/47	medium · avg 36%	▇ 2.3%	sharp · at its level · noisy
LiveCodeBench v2	71.2% ✎	▄··█·▄···▄▄█ beats 6/7	medium · avg 59%	▇ 2.0%	at its level · small run
CodeReviewQA: automated code refinement, exact match rate	30.6% ✎	▁▁▁▁▁▁▂▂▃█▅▄ beats 21/47	hard · avg 33%	▇ 1.5%	at its level
CodeReviewQA: solution identification (hard), invariant accuracy	76.4% ✎	▃··▁▂▂▃▂█▃▄▂ beats 39/47	medium · avg 54%	▇ 1.5%	sharp · at its level · noisy
CodeReviewQA: change type recognition, invariant accuracy	73.2% ✎	▁▁▁······▁▃█ beats 24/47	easy · avg 67%	▇ 0.0%	blunt · at its level · noisy

math▁▂▃▅▆▅▄█▄▁▁▁ #171 of 398 · 39% of its score · 7 tests

Beats MiniMax-M2.1, Mistral Medium 3.5 (128B), GPT-4.1 here, though they rank higher overall. Trails OpenThinker3 7B, Olmo 3 7B Think, OpenThinker2 32B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2024 lab +2	72.6% ✎	▃·▃·▃▃···█·▃ beats 6/7	medium · avg 56%	▇▇▇▇▇ 14%	sharp · at its level · several runs
MATH-500 lab +2	94.3% ✎	▄·▄▄·▄▄█···▄ beats 5/7	very easy · avg 86%	▇▇▇▇ 9.1%	several runs
AIME 2025	50.0% ✎	▁·▁▁▁▂▅▆█▇▅▁ beats 36/133	medium · avg 38%	▇▇▇ 6.6%	sharp · at its level
OTIS mock AIME 2024–2025	55.6% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 83/193	very hard · avg 8.5%	▇▇ 4.9%	sharp · noisy
OMEGA (math generalisation)	38.9% ✎	▅···▃▃···▃▃█ beats 4/8	hard · avg 24%	▇ 2.1%	at its level
THINK-Bench final-answer accuracy (reasoning models)	75.9% ✎	▄█·▄····▄▄·█ beats 3/7	easy · avg 76%	▇ 1.4%	typical
MATH (subset unspecified)	92.6% ✎	▅····▃▃··▃▃█ beats 4/8	easy · avg 74%	▇ 1.0%	noisy

other tests7 not in a task · 41% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond +3	64.1% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 93/208	hard · avg 40%	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 34%	sharp · at its level · several runs
BRIDGE clinical text, zero-shot average (87 tasks)	39.8% ✎	▁▁▁▁▁▂▂▃▄▆█▃ beats 69/97	hard · avg 32%	▇ 3.5%	blunt · at its level
SurgiQ surgical MCQ (zero-shot, log-probability scoring)	60.4% ✎	▂···▂▂·▃▂▃██ beats 12/17	medium · avg 53%	▇ 1.6%	at its level
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy	79.5% ✎	▃·▁▁▂▁▂▅█▇▄▁ beats 70/100	medium · avg 68%	▇ 1.0%	at its level · noisy
Chess Puzzles (Epoch)	1.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 27/135	very hard · avg 5.9%	▇ 0.6%	sharp · too hard for it · noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy	74.5% ✎	▁▁▁▁▁▁▁▂▄█▄▃ beats 48/146	easy · avg 83%	▇ 0.5%	noisy
Open PT-LLM Leaderboard average (9 Portuguese tasks)	26.2% ✎	▃▁▁▁▁▁▁▁▂█▅▁ beats 0/276	medium · avg 73%	—	failed run: Failed run: several task scores fall far below the trivial baselines (OAB Exams 5.06 on 4-option questions, PT Hate Speech 3.73 and tweetSentBR 3.42 macro-F1) while ASSIN2 RTE is 89.84, i.e. most answers in those tasks were not parsed (reasoning output); the 26.15 average (random baseline 32.64) is not a measurement of ability.

$ where to run it

Through an API

api · Cloudflare · $1.59/M · $0.497 in / $4.881 out per 1M tokens · 131k context · developers.cloudflare.com ↗ · price sourceWorkers AI published marginal USD per 1M tokens after free allocation. Workers plan fees excluded; billed in Neurons. Published USD rates are rounded. Product: Workers AI. API model: @cf/deepseek-ai/deepseek-r1-distill-qwen-32b. Settings: plan=Workers Paid, free allocation=excluded, billing=published marginal USD rates; service bills Neurons. Currency: USD.

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run DeepSeek R1 Distill Qwen 32B · needs ~19.9 GB · from huggingface.co/lmstudio-community
chat in the terminal
llama cli -hf lmstudio-community/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf lmstudio-community/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run DeepSeek R1 Distill Qwen 32B · needs ~19.9 GB · from huggingface.co/lmstudio-community
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/lmstudio-community/DeepSeek-R1-Distill-Qwen-32B-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run DeepSeek R1 Distill Qwen 32B · needs ~20.18 GB · from ollama.com
chat in the terminal
ollama run deepseek-r1:32b

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run DeepSeek R1 Distill Qwen 32B · needs ~18.4 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit

Weights and documentation: huggingface.co/deepseek-ai ↗

$ how the score was worked out · open-thoughts/OpenThinker2-32… 26%, deepseek-ai/DeepSeek-R1 model… 17%, open-thoughts/OpenThinker-32B… 17% and 8 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
open-thoughts/OpenThinker2-32B model card	4	▇▇▇▇▇▇▇▇ 26%	Apache-2.0
deepseek-ai/DeepSeek-R1 model card · lab-reported	4	▇▇▇▇▇ 17%	mit
open-thoughts/OpenThinker-32B model card	4	▇▇▇▇▇ 17%	Apache-2.0
Epoch AI — internal evaluations	3	▇▇▇▇▇ 16%	CC BY 4.0
WebApp1K models leaderboard (onekq-ai)	2	▇▇ 6%	MIT
CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models (arXiv:2503.16167v2)	4	▇▇ 5%	CC-BY-4.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)	1	▇ 3%	MIT
allenai/Olmo-3.1-32B-Think model card	2	▇ 3%	Apache-2.0
SurgiQ: A Large-Scale Multi-Domain Benchmark for Evaluating Surgical Understanding in Large Language Models (arXiv:2606.08071v1)	1	▇ 2%	CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇ 1%	MIT
THINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models (arXiv:2505.22113v1)	1	▇ 1%	CC-BY-4.0
prior · pull toward the middle		▇ 0%
weighted average		= 55.9

not counted (1)

Open Portuguese LLM Leaderboard (archived)

Open PT-LLM Leaderboard average (9 Portuguese tasks) [correct]

26.2%

failed run: Failed run: several task scores fall far below the trivial baselines (OAB Exams 5.06 on 4-option questions, PT Hate Speech 3.73 and tweetSentBR 3.42 macro-F1) while ASSIN2 RTE is 89.84, i.e. most answers in those tasks were not parsed (reasoning output); the 26.15 average (random baseline 32.64) is not a measurement of ability.

$ quantization evidence · 33 measurements

How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent.

precision	test	full → quantized	kept
AWQ 4-bit (AutoAWQ; vLLM)	AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.644 → 0.678	105.3%	source ↗
AWQ 4-bit (AutoAWQ; vLLM)	Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.822 → 0.831	101.1%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	MATH-500 pass@1	0.951 → 0.953	100.2%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)	0.827 → 0.826	99.9%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	OpenLLM v2 average of 6 tasks (IFEval, BBH, MATH-Hard, GPQA, MuSR, MMLU-Pro; lm-evaluation-harness `leaderboard`, chat template, vLLM; values exactly as published, per-task normalisation not stated)	0.369 → 0.369	99.8%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	MMLU 5-shot acc (lm-evaluation-harness, vLLM)	0.811 → 0.81	99.8%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)	0.745 → 0.744	99.8%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	MMLU-Pro 5-shot (OpenLLM v2 config; value as published, normalisation not stated)	0.498 → 0.494	99.2%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)	0.763 → 0.755	99.0%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	AIME 2024 pass@1	0.698 → 0.685	98.2%	source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)	GPQA Diamond pass@1	0.64 → 0.629	98.2%	source ↗
GPTQ 3-bit (vLLM)	Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.822 → 0.725	88.2%	source ↗
GPTQ 3-bit (vLLM)	AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.644 → 0.444	68.9%	source ↗
GPTQ 4-bit (vLLM)	AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.644 → 0.689	107.0%	source ↗
GPTQ 4-bit (vLLM)	Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)	0.822 → 0.83	101.0%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)	0.827 → 0.844	102.0%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	MATH-500 pass@1	0.951 → 0.95	99.9%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)	0.745 → 0.743	99.7%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	MMLU 5-shot acc (lm-evaluation-harness, vLLM)	0.811 → 0.809	99.7%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	GPQA Diamond pass@1	0.64 → 0.638	99.6%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)	0.763 → 0.759	99.4%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	AIME 2024 pass@1	0.698 → 0.688	98.7%	source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)	MMLU-Pro 5-shot (OpenLLM v2 config; value as published, normalisation not stated)	0.498 → 0.484	97.2%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)	0.827 → 0.839	101.4%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	GPQA Diamond pass@1	0.64 → 0.647	101.1%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	OpenLLM v2 average of 6 tasks (IFEval, BBH, MATH-Hard, GPQA, MuSR, MMLU-Pro; lm-evaluation-harness `leaderboard`, chat template, vLLM; values exactly as published, per-task normalisation not stated)	0.369 → 0.37	100.2%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)	0.745 → 0.745	100.0%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	MATH-500 pass@1	0.951 → 0.95	99.9%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	MMLU 5-shot acc (lm-evaluation-harness, vLLM)	0.811 → 0.809	99.8%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)	0.763 → 0.76	99.6%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	MMLU-Pro 5-shot (OpenLLM v2 config; value as published, normalisation not stated)	0.498 → 0.491	98.6%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	IFEval 0-shot instruction-level strict acc (OpenLLM v2 config, chat template)	0.429 → 0.419	97.8%	source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)	AIME 2024 pass@1	0.698 → 0.682	97.7%	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json