[whatmodel.sh](../../README.md) / DeepSeek V3.2 # DeepSeek V3.2 DeepSeek · released 2025-12 · 685B params (37B active) · mit **score 67.1** ±0.8 · #112 of 732 · from 30 benchmark results across 21 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at ● marks DeepSeek V3.2 among every model on that task or test, worst to best. 
task            vs the others         rank  counts  tests
agent tasks     ▁▁▃▆█▄▅●▄▄▂▁   #110 of 313     14%      8
coding          ▁▁▁▃▅▇█▇●▄▃▁   #121 of 469     16%      4
chat & writing  ▁▁▁▃▆▄▆█▆●▇▃   #90 of 402*    6.0%      1
math            ▁▁▃▅█▆▄●▄▂▁▁    #85 of 431     21%      7
agent tasks · #110 of 313 · 14% of its score Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [o3-pro](o3-pro.md), [Grok 4.3](grok-4-3.md) here, though they rank higher overall. Trails [SmolLM3 3B](smollm3-3b.md), [Nova 2 Pro (Preview)](nova-2-pro.md), [Mistral Medium 3.5 (128B)](mistral-medium-3-5.md) here, though they rank lower overall. 
test                                result  vs the others in its run   difficulty            counts  why
MCP-Atlas (public set)               62.2%  ▄······●▄▄▄█ beats 1/6     hard · avg 26%          7.7%  at its level · several runs
τ²-bench aggregate                   85.2%  ▄······▄●·▄▄ beats 2/5     medium · avg 51%        2.8%  typical
Vibe Code Bench (test split, 50 w…    5.1%  ▃··●·▃██▃▃▃█ beats 1/14    very hard · avg 2.8%    2.0%  sharp · noisy
EEGAgentBench overall (tool-based…   46.0%  ▂·▁▁··▇●█▂·▁ beats 12/25   hard · avg 28%          1.8%  blunt · at its level
SWE-bench Verified (maker)           73.1%  ▁▁▁▁▁▁▂▄●▆▁▁ beats 93/138  hard · avg 19%             —  not counted yet: no other model in the …
τ²-bench Telecom (maker)             96.2%  ▁·▅█▆▅▂▃▃●▃▁ beats 76/89   very hard · avg 8.8%       —  not counted yet: no other model in the …
Terminal-Bench 2.0 (maker)           46.4%  ▂▂▅▅██●█▆▂▃▃ beats 18/36   very hard · avg 10%        —  not counted yet: no other model in the …
Toolathlon (Tool Decathlon) (make…   35.2%  ▂▂▄▂▄●▆▂█▆██ beats 7/27                               —  not counted yet: no other model in the …
coding · #121 of 469 · 16% of its score Beats [Qwen3.5 Plus](qwen3-5-plus.md) here, though they rank higher overall. Trails [Llama 3.1 Nemotron Ultra 253B v1](llama-3-1-nemotron-ultra-253b.md), [Solar Pro 4](solar-pro-4.md), [Laguna XS 2.1 (33B-A3B)](laguna-xs-2-1.md) here, though they rank lower overall. 
test                                result  vs the others in its run   difficulty            counts  why
Arena WebDev (Code Arena)           1362.5  ▁▁▃▂●█▃▆▅▅▁▂ beats 34/129  avg 981                  11%  sharp
WebApp1K-Duo-React pass@1 (onekq-…   65.0%  ▁··▁·▁▁▁▆●▇█ beats 20/45   hard · avg 29%          2.9%  at its level
Vibe Code Bench (test split, 50 w…    5.1%  ▃··●·▃██▃▃▃█ beats 1/14    very hard · avg 2.8%    2.0%  sharp · noisy
SWE-bench Verified (maker)           73.1%  ▁▁▁▁▁▁▂▄●▆▁▁ beats 93/138  hard · avg 19%             —  not counted yet: no other model in the …
chat & writing · #90 of 402 · 6.0% of its score Beats [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md), [Inkling Small (266B-A12B)](inkling-small.md), [Qwen3.5 122B-A10B](qwen3-5-122b-a10b.md) here, though they rank higher overall. Trails [Gemma 4 12B](gemma-4-12b.md), [Gemma 4 26B-A4B](gemma-4-26b-a4b.md), [Mistral Medium 3.5 (128B)](mistral-medium-3-5.md) here, though they rank lower overall. 
test            result  vs the others in its run    difficulty  counts  why
LMArena (text)  1422.3  ▁▁▁▂▃▃▃▆▆●█▃ beats 275/385  avg 1260      6.0%  blunt
math · #85 of 431 · 21% of its score Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [Claude Opus 4.5](claude-opus-4-5.md), [Qwen3.5 397B-A17B](qwen3-5-397b-a17b.md) here, though they rank higher overall. Trails [Gemini 2.5 Flash](gemini-2-5-flash.md), [Granite 4.2 30B](granite-4-2-30b.md), [Qwen3-Next 80B-A3B Instruct](qwen3-next-80b-a3b-instruct.md) here, though they rank lower overall. 
test                                result  vs the others in its run    difficulty            counts  why
MathArena ArXivMath January 2026 …   57.1%  ▂··▄▂▄█●▆▄▄█ beats 11/23    hard · avg 23%          5.1%  at its level
AIME 2025 (maker)                    93.1%  ▁·▁▁▁▂▅▆█▇●▁ beats 127/133  medium · avg 53%        3.7%  sharp · small run
MathArena ArXivMath December 2025…   41.5%  ▃·▃▃●▁·▄·▁▁▁ beats 9/18     hard · avg 19%          3.4%  at its level
OTIS mock AIME 2024–2025             87.8%  ▃▃▄▃▃▃▇▇●▃▂▅ beats 137/197  very hard · avg 7.2%    3.1%  sharp · noisy
HMMT 2025 (edition unspecified) (…   90.2%  ▃█▇▅▇▅●·▁··▁ beats 30/32    medium · avg 48%        3.0%  sharp · small run
MathNet-Solve-Test micro-average …   40.1%  ▃▃···●▃·▃·▃▃ beats 2/8      very hard · avg 15%     2.0%  at its level
FormalProofBench (200 graduate-le…    3.0%  ●▅·▃··▅▃█··▃ beats 0/10     very hard · avg 3.1%    0.9%  noisy
other tests (15) 
test                                result  vs the others in its run    difficulty            counts  why
GPQA Diamond                         83.4%  ▁▁▁▁▂▄▆▄▅●█▅ beats 139/211  hard · avg 39%          8.0%  sharp · at its level
ARC-AGI-1 public evaluation set (…   61.6%  ▁·▇▃▁·●▇▁·▇▃ beats 13/27    very hard · avg 4.0%    4.4%  sharp · at its level · noisy
DSEffi-Bench Pass (1,000 data-sci…   56.1%  ▅··▃▅·█·▃●█▃ beats 9/15     hard · avg 23%          4.4%  at its level
KINA (Knowledge Index of Noah's A…   38.0%  ▁▁▂▁▃▄█▆▄●▂▃ beats 30/39    very hard · avg 21%     4.4%  at its level
ARC-AGI-2 public evaluation set (…    3.9%  █··●▅▁▃▁▃·▁▃ beats 8/24     very hard · avg 0.3%    3.3%  sharp · noisy
The Token Games Elo (10-model puz…  1007.0  ●··▄█··▄·▄·▄ beats 1/7      avg 822                 3.2%  typical
WildClawBench                        34.0%  ▃·▅▅●·▅█▅▃▃▃ beats 6/17     very hard · avg 11%     2.8%  at its level
Chess Puzzles (Epoch)                14.0%  █▁▁▁▁●▃▃▃▂▁▁ beats 65/137   very hard · avg 5.8%    2.8%  sharp · noisy
CPTUB average (SpeakLeash, 0-5 ju…   82.8%  ▁▂▂▂▃█▃▆▅▄●▃ beats 87/95    medium · avg 65%        2.3%  typical
ResearchRubrics                      55.8%  █●█··██·█··█ beats 1/6      hard · avg 33%          2.3%  at its level
RPC-Bench overall, text input (Z.…   32.0%  ▄··▄▄·▄▄●·▄▄ beats 5/8      very hard · avg 12%     2.3%  at its level
Humanity's Last Exam (maker)         25.1%  ▁▁▁▄▆▆▇●█▆▁▃ beats 63/111   very hard · avg 4.7%    2.0%  small run
HLE-Verified full set, text-only …   36.4%  ●·█·····█··█ beats 0/3      very hard · avg 15%     1.7%  at its level
GAIA (no-file subset)                75.1%  ▃▃····●▃···▃ beats 2/6      medium · avg 49%        1.2%  at its level
Vectara hallucination leaderboard…   93.7%  ▁▂▁▂▅█▂▃●▃▁▁ beats 75/98    very easy · avg 90%     0.0%  blunt
## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's 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 is added up by who ran the tests. [Method](../method.md) 
source                                    results  share of the score
LMArena leaderboard dataset                     2   17% ▇▇
Epoch AI                                        3   14% ▇▇
deepseek-ai/DeepSeek-V3.2 model card (s…        3    9% ▇
MathArena ArXivMath / BrokenArXiv / ArX…        2    8% ▇
ARC Prize Foundation model_baseline out…        2    8% ▇
stepfun-ai/Step-3.5-Flash model card            3    6% ▇
DSEffi-Bench: Demystifying Large Langua…        1    4% ▇
Knowledge Index of Noah's Ark                   1    4% ▇
zai-org/GLM-5 model card                        1    4% ▇
GLM-5 Team                                      1    4% ▇
The Token Games: Evaluating Language Mo…        1    3% ▇
WebApp1K models leaderboard                     1    3% ▇
WildClawBench: A Benchmark for Real-Wor…        1    3% ▇
SpeakLeash CPTUB                                1    2% ▇
RPC-Bench leaderboard                           1    2% ▇
Vibe Code Bench                                 1    2% ▇
MathNet: a Global Multimodal Benchmark …        1    2% ▇
EEGAgentBench: Benchmarking LLM Agents …        1    2% ▇
HLE-Verified: A Systematic Verification…        1    2% ▇
FormalProofBench                                1    1% ▇
Vectara hallucination leaderboard               1    0% ▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 67.1
Licences: LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · deepseek-ai/DeepSeek-V3.2 model card ([MIT](../../licenses/deepseek-ai--DeepSeek-V3.2.txt)) · MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets ([CC BY-SA 4.0](../../licenses/matharena-sharealike-NOTICE.txt)) · ARC Prize Foundation model\_baseline outputs on ARC-AGI-1/2 public evaluation sets ([MIT](../../licenses/arcprize-model-baseline-outputs-MIT.txt)) · stepfun-ai/Step-3.5-Flash model card ([Apache-2.0](../../licenses/stepfun-ai--Step-3.5-Flash.txt)) · DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation ([CC BY 4.0](../../licenses/paper-2608.30248v1-CC-BY-4.0.txt)) · Knowledge Index of Noah's Ark ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · zai-org/GLM-5 model card ([MIT](../../licenses/zai-org--GLM-5.txt)) · GLM-5 Team ([CC BY 4.0](../../licenses/report-2602.15763v2.txt)) · The Token Games: Evaluating Language Model Reasoning with Puzzle Duels ([CC BY-SA 4.0](../../licenses/paper-2602.17831v2-CC-BY-SA-4.0.txt)) · WebApp1K models leaderboard ([MIT](../../licenses/webapp1k-leaderboard-MIT.txt)) · WildClawBench: A Benchmark for Real-World ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · SpeakLeash CPTUB ([MIT](../../licenses/hf-leaderboard-spaces-MIT.txt)) · RPC-Bench leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · Vibe Code Bench ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · MathNet: a Global Multimodal Benchmark for Mathematical Reasoning and Retrieval ([CC BY 4.0](../../licenses/paper-2604.18584v2-CC-BY-4.0.txt)) · EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam ([CC BY 4.0](../../licenses/paper-2602.13964v4-CC-BY-4.0.txt)) · FormalProofBench ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · Vectara hallucination leaderboard ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0))
Not counted (5) 
source                  test                         score  why
deepseek-ai/DeepSeek-…  SWE-bench Verified           73.1%  too few models have it yet
deepseek-ai/DeepSeek-…  τ²-bench aggregate           80.3%  only model in this source's run
deepseek-ai/DeepSeek-…  Terminal-Bench 2.0           46.4%  too few models have it yet
deepseek-ai/DeepSeek-…  Toolathlon (Tool Decathlon)  35.2%  too few models have it yet
deepseek-ai/DeepSeek-…  τ²-bench Telecom             96.2%  too few models have it yet
## How to run it ### Through an API api (Deep Infra) · $0.29/M blended · $0.26 in / $0.38 out per 1M tokens · [deepinfra.com](https://deepinfra.com/models) [price source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/providers/deepinfra/models/deepseek-ai/DeepSeek-V3.2.toml): Standard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/deepseek/deepseek-v3.2.toml) ### llama.cpp (recommended) 1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) ```sh # macOS / Linux curl -LsSf https://llama.app/install.sh | sh # Windows irm https://llama.app/install.ps1 | iex ``` 2. Run DeepSeek V3.2 (ud-q4_k_xl, needs ~411.5 GB, from [huggingface.co/unsloth](https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF/tree/main)) ```sh llama cli -hf unsloth/DeepSeek-V3.2-GGUF:UD-Q4_K_XL ``` 3. Or start a local server and open the address it prints (http://localhost:9931) ```sh llama serve -hf unsloth/DeepSeek-V3.2-GGUF:UD-Q4_K_XL ``` Other versions: ud-q2_k_xl (~206 GB) `llama cli -hf unsloth/DeepSeek-V3.2-GGUF:UD-Q2_K_XL` · ud-iq1_s (~103.25 GB) `llama cli -hf unsloth/DeepSeek-V3.2-GGUF:UD-IQ1_S` ### LM Studio 1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download) 2. Run DeepSeek V3.2 (ud-q4_k_xl, needs ~411.5 GB, from [huggingface.co/unsloth](https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF/tree/main)) ```sh lms get https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF@UD-Q4_K_XL -y ``` 3. Then run lms ls to see its name, and lms chat followed by that name Other versions: ud-q2_k_xl (~206 GB) `lms get https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF@UD-Q2_K_XL -y` · ud-iq1_s (~103.25 GB) `lms get https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF@UD-IQ1_S -y` · 5bit (~462.1 GB, Mac only) `lms get https://huggingface.co/mlx-community/DeepSeek-V3.2-mlx-5bit -y` ### MLX (Mac) 1. Install MLX (Mac). Apple silicon Macs only, text only. ```sh # with uv uv tool install mlx-lm # or pip install mlx-lm ``` 2. Run DeepSeek V3.2 (5bit, needs ~462.1 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/DeepSeek-V3.2-mlx-5bit/tree/24cc22660535230a3a69d0742547519819aa1059)) ```sh mlx_lm.chat --model mlx-community/DeepSeek-V3.2-mlx-5bit ``` 3. Or start an OpenAI-compatible server at http://localhost:8080/v1 ```sh mlx_lm.server --model mlx-community/DeepSeek-V3.2-mlx-5bit ``` Model repository and documentation: [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-V3.2) [interactive version](https://whatmodel.sh/model/deepseek-v3-2) · [JSON](https://whatmodel.sh/api?format=json&all=1)