whatmodel.sh / DeepSeek V3.2

DeepSeek V3.2

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

DeepSeek V3.2 is an open-weights language model from DeepSeek, released December 2025. Its published size is 685B parameters (37B active per token) and a 128k-token context window. Licence: mit.

68.1±0.8#104 of 719 LLMsfrom 30 benchmark results across 21 sources

$ how it compares

The LLMs ranked either side of DeepSeek V3.2, out of 828 with a score, then more from DeepSeek.

#modelscorewhereRAMreleased
114LongCat-Flash-Thinking (560B-A27B) Meituan69±1comparison only2025-09
115Step 3.5 Flash (196B-A11B) StepFun68±1api $0.15/M +local~118GB2026-02
116DeepSeek V3.2 DeepSeek68±1api $0.29/M +local~110GB2025-12
117MiMo-V2-Omni Xiaomi68±2comparison only2026-03
118Gemini 2.5 Pro Experimental (03-25) Google68±3comparison only2025-03
more from DeepSeek
21DeepSeek V4.1 Flash DeepSeek78±1api $0.26/M +local~537GB2026-09
31DeepSeek V4 Pro (0813) DeepSeek76±1api $1.63/M +local~1010GB2026-08
35DeepSeek V3.2 Speciale DeepSeek76±3comparison only

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #99 of 300 · 15% of its score · 8 tests

Beats Grok 4.20 (0309 reasoning), Grok 4.3, o3-pro here, though they rank higher overall. Trails SmolLM3 3B, Nova 2 Pro (Preview), MiniMax-M2.1 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
MCP-Atlas (public set) +162.2% ✎▄······▄▄▄▄█ beats 1/6hard · avg 25%▇▇▇ 7.9%at its level · several runs
τ²-bench aggregate +185.2% ✎▄······▄█·▄▄ beats 2/5medium · avg 46%▇ 3.5%typical
Vibe Code Bench (test split, 50 web apps)5.1% ✎▃··▃·▃██▃▃▃█ beats 1/14very hard · avg 2.3%▇ 2.4%sharp · noisy
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)46.0% ✎▂·▁▁··▇▇█▂·▁ beats 12/25hard · avg 29%▇ 1.6%blunt · at its level
SWE-bench Verified lab73.1% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 88/131hard · avg 19%—not counted yet: no other model in the same run to compare it with
τ²-bench Telecom lab96.2% ✎▁·▅█▆▅▂▃▃▃▃▁ beats 76/89very hard · avg 9.6%—not counted yet: no other model in the same run to compare it with
Terminal-Bench 2.0 lab46.4% ✎▂▂▅▅████▆▂▃▃ beats 18/36very hard · avg 9.5%—not counted yet: no other model in the same run to compare it with
Toolathlon (Tool Decathlon) lab35.2% ✎▂▂▄▂▄▄▆▂█▆██ beats 7/27—not counted yet: no other model in the same run to compare it with
coding▁▁▁▃▆██▇▆▄▃▁ #112 of 458 · 15% of its score · 4 tests

Beats Qwen3.5 Plus here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4, Laguna XS 2.1 (33B-A3B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1361.7 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 34/119avg 981▇▇▇▇ 10%sharp
Vibe Code Bench (test split, 50 web apps)5.1% ✎▃··▃·▃██▃▃▃█ beats 1/14very hard · avg 2.3%▇ 2.4%sharp · noisy
WebApp1K-Duo-React pass@1 (onekq-ai)65.0% ✎▁··▁·▁▁▁▆█▇█ beats 20/45hard · avg 32%▇ 2.4%at its level
SWE-bench Verified lab73.1% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 88/131hard · avg 19%—not counted yet: no other model in the same run to compare it with
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #85 of 393 * · 5.5% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), Inkling Small (266B-A12B), Qwen3.5 122B-A10B here, though they rank higher overall. Trails Gemma 4 12B, Gemma 4 26B-A4B, Mistral Medium 3.5 (128B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1422.7 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 273/379avg 1268▇▇ 5.5%blunt
math▁▂▃▅▆▅▄█▄▁▁▁ #75 of 398 · 22% of its score · 7 tests

Beats Claude Sonnet 4.6, Claude Opus 4.5, Qwen3.7 Plus here, though they rank higher overall. Trails Granite 4.2 30B, INTELLECT-3, EXAONE 4.5 33B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
MathArena ArXivMath January 2026 (23 problems)57.1% ✎▂··▄▂▄█▆▆▄▄█ beats 11/23hard · avg 22%▇▇ 5.2%at its level
AIME 2025 lab93.1% ✎▁·▁▁▁▂▅▆█▇▅▁ beats 127/133medium · avg 52%▇▇ 3.9%sharp · small run
HMMT 2025 (edition unspecified) lab90.2% ✎▃█▇▅▇▅▃·▁··▁ beats 30/32medium · avg 43%▇ 3.6%sharp · small run
OTIS mock AIME 2024–202587.8% ✎▃▃▄▃▃▃▇▇█▃▂▅ beats 137/193very hard · avg 8.5%▇ 3.5%sharp · noisy
MathArena ArXivMath December 2025 (17 problems)41.5% ✎▃·▃▃█▁·▄·▁▁▁ beats 9/18hard · avg 19%▇ 3.3%at its level
MathNet-Solve-Test micro-average accuracy (6,400 olympiad problems)40.1% ✎▃▃···█▃·▃·▃▃ beats 2/8hard · avg 15%▇ 2.0%at its level
FormalProofBench (200 graduate-level Lean 4 problems, agentic)3.0% ✎▃▅·▃··▅▃█··▃ beats 0/10very hard · avg 3.0%▇ 0.9%noisy
other tests15 not in a task · 44% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond83.4% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 139/208hard · avg 40%▇▇▇ 7.8%sharp · at its level
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)61.6% ✎▁·▇▃▁·█▇▁·▇▃ beats 13/27very hard · avg 4.1%▇▇ 4.5%sharp · at its level · noisy
DSEffi-Bench Pass (1,000 data-science tasks, greedy)56.1% ✎▅··▃▅·█·▃██▃ beats 9/15hard · avg 24%▇▇ 4.0%at its level
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items38.0% ✎▁▁▂▁▃▄█▆▄▅▂▃ beats 30/39very hard · avg 22%▇▇ 3.8%blunt · at its level
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)3.9% ✎█··▃▅▁▃▁▃·▁▃ beats 8/24very hard · avg 0.3%▇ 3.5%sharp · noisy
The Token Games Elo (10-model puzzle-duel tournament)1007.0 ✎█··▄█··▄·▄·▄ beats 1/7avg 815▇ 3.2%typical
WildClawBench34.0% ✎▃·▅▅█·▅█▅▃▃▃ beats 6/17very hard · avg 11%▇ 2.9%at its level
ResearchRubrics55.8% ✎███··██·█··█ beats 1/6hard · avg 31%▇ 2.6%at its level
7 more tests
testresultvs the others in its rundifficultycountswhy
Chess Puzzles (Epoch)14.0% ✎█▁▁▁▁▃▁▄▂▂▁▁ beats 65/135very hard · avg 5.9%▇ 2.5%sharp · noisy
RPC-Bench overall, text input (Z.ai)32.0% ✎▄··▄▄·▄▄█·▄▄ beats 5/8very hard · avg 11%▇ 2.4%at its level
Humanity's Last Exam lab25.1% ✎▁▁▁▃▆▆▇██▆▁▃ beats 59/104very hard · avg 4.7%▇ 2.0%small run
CPTUB average (SpeakLeash, 0-5 judge scale)82.8% ✎▁▂▂▂▃█▃▆▅▄▃▃ beats 87/95easy · avg 66%▇ 2.0%typical
HLE-Verified full set, text-only (Qwen Team runs)36.4% ✎█·█·····█··█ beats 0/3very hard · avg 14%▇ 1.8%at its level
GAIA (no-file subset)75.1% ✎▃▃····█▃···▃ beats 2/6medium · avg 47%▇ 1.3%at its level
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)93.7% ✎▁▂▁▂▅█▂▃▄▃▁▁ beats 75/98very easy · avg 90%▇ 0.0%blunt

$ where to run it

Through an API

api · Deep Infra · $0.29/M · $0.26 in / $0.38 out per 1M tokens · 128k context · deepinfra.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

version

  1. 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 ↗

  2. Run DeepSeek V3.2 · needs ~411.5 GB · from huggingface.co/unsloth
    chat in the terminal
    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)
    llama serve -hf unsloth/DeepSeek-V3.2-GGUF:UD-Q4_K_XL

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run DeepSeek V3.2 · needs ~411.5 GB · from huggingface.co/unsloth

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/unsloth/DeepSeek-V3.2-GGUF@UD-Q4_K_XL -y
  4. Then run lms ls to see its name, and lms chat followed by that name

MLX (Mac)

  1. Install MLX (Mac) · Apple silicon Macs only, text only.
    with uv
    uv tool install mlx-lm
    or
    pip install mlx-lm
  2. Run DeepSeek V3.2 · needs ~462.1 GB, Mac only · from huggingface.co/mlx-community
    chat in the terminal
    mlx_lm.chat --model mlx-community/DeepSeek-V3.2-mlx-5bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model mlx-community/DeepSeek-V3.2-mlx-5bit

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

$ how the score was worked out · LMArena leaderboard dataset 15%, Epoch AI — internal evaluatio… 14%, deepseek-ai/DeepSeek-V3.2 mod… 10% and 18 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

sourceresultsshare of the scorelicence
LMArena leaderboard dataset2▇▇▇▇▇ 15%CC BY 4.0
Epoch AI — internal evaluations3▇▇▇▇ 14%CC BY 4.0
deepseek-ai/DeepSeek-V3.2 model card · lab-reported3▇▇▇ 10%MIT
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets2▇▇▇ 9%CC BY-SA 4.0
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets2▇▇ 8%MIT
stepfun-ai/Step-3.5-Flash model card3▇▇ 7%Apache-2.0
DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation — authors' own model runs1▇ 4%CC BY 4.0
zai-org/GLM-5 model card1▇ 4%MIT
GLM-5 Team (Zhipu AI) — GLM-5 technical report1▇ 4%CC BY 4.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)1▇ 4%CC BY 4.0
The Token Games: Evaluating Language Model Reasoning with Puzzle Duels — authors' own tournament1▇ 3%CC BY-SA 4.0
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation (arXiv:2605.10912v1)1▇ 3%CC BY 4.0
Vibe Code Bench (Vals AI, arXiv:2603.04601v3)1▇ 2%CC-BY-4.0
WebApp1K models leaderboard (onekq-ai)1▇ 2%MIT
RPC-Bench leaderboard (Z.ai research paper comprehension)1▇ 2%Apache-2.0
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)1▇ 2%MIT
MathNet: a Global Multimodal Benchmark for Mathematical Reasoning and Retrieval — authors' own model runs1▇ 2%CC BY 4.0
HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam — Qwen Team re-evaluation on revised Humanity's Last Exam1▇ 2%CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)1▇ 2%CC BY 4.0
FormalProofBench (Vals AI, arXiv:2603.26996v1)1▇ 1%CC-BY-4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)1▇ 0%Apache-2.0
prior · pull toward the middle▇ 0%
weighted average= 68.1

not counted (5)

deepseek-ai/DeepSeek-V3.2 model cardSWE-bench Verified [correct]73.1%fewer than three models have this test yet, or this source ran it for this model alone
deepseek-ai/DeepSeek-V3.2 model cardτ²-bench aggregate [correct]80.3%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
deepseek-ai/DeepSeek-V3.2 model cardTerminal-Bench 2.0 [correct]46.4%fewer than three models have this test yet, or this source ran it for this model alone
deepseek-ai/DeepSeek-V3.2 model cardToolathlon (Tool Decathlon) [correct]35.2%fewer than three models have this test yet, or this source ran it for this model alone
deepseek-ai/DeepSeek-V3.2 model cardτ²-bench Telecom [correct]96.2%fewer than three models have this test yet, or this source ran it for this model alone

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit 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 · sources and licences · json