whatmodel.sh / DeepSeek V3.2 Speciale

DeepSeek V3.2 Speciale

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

DeepSeek V3.2 Speciale is an open-weights language model from DeepSeek. Its published size is 685B parameters (37B active per token). Licence: mit.

Catalogue identity notes

Comparison-only. DeepSeek's high-compute reasoning variant of V3.2 (repo deepseek-ai/DeepSeek-V3.2-Speciale @c562883, MIT; card: 'designed exclusively for deep reasoning tasks and does not support the tool-calling functionality'). Distinct from deepseek-v3-2 (the hatched 'DeepSeek-V3.2-Thinking' bars). Added for the Speciale bars of DeepSeek's own V3.2 chart. No verified run route recorded.

Model information sources

· Parameter count ↗

685,396,921,376 parameters in the repo's safetensors metadata; DeepSeek-V3.2 card: 'The model structure of DeepSeek-V3.2 and DeepSeek-V3.2-Speciale are the same as DeepSeek-V3.2-Exp' (37B active, as catalogued for deepseek-v3-2).

75.8±2.8comparison only: no current run option verifiedfrom 3 benchmark results across 1 source

$ how it compares

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

#modelscorewhereRAMreleased
33GLM-5.3 Z.ai76±1api $2.15/M +local~239GB2026-08
34Qwen3.8 Max (0902) Alibaba76±1api $3.00/M2026-09
35DeepSeek V3.2 Speciale DeepSeek76±3comparison only
36DeepSeek V4 Flash Vision (Exp) DeepSeek76±2local~104GB2026-08
37Qwen3.8-Flash-Next Alibaba75±1local~59GB2026-08
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
50DeepSeek V4 Flash (0731) DeepSeek74±1api $0.11/M +local~92GB2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

math▁▂▃▅▇▅▅█▄▃▁▁ #20 of 462 * · 67% of its score · 2 tests

Beats DeepSeek V4.1 Flash, Claude Haiku 5.5, Gemini 3.8 Flash here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
AIME 2025 lab96.0% ✎▁·▁▁▁▂▅▆█▇▅▃ beats 131/133medium · avg 52%▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 35%sharp · maxed out · small run
HMMT 2025 (edition unspecified) lab99.2% ✎▃█▇▅▇▅▁·▁··▃ beats 32/32medium · avg 43%▇▇▇▇▇▇▇▇▇▇▇▇▇ 32%sharp · maxed out · small run
other tests1 not in a task · 30% of its score
testresultvs the others in its rundifficultycountswhy
Humanity's Last Exam lab30.6% ✎▁▁▁▃▆▆▇██▆▁▃ beats 73/104very hard · avg 4.7%▇▇▇▇▇▇▇▇▇▇▇▇ 30%at its level · small run

$ where to run it

No verified API or download recorded yet · add one

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

$ how the score was worked out · deepseek-ai/DeepSeek-V3.2 mod… 97%

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
deepseek-ai/DeepSeek-V3.2 model card · lab-reported3▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 97%MIT
prior · pull toward the middle▇ 3%
weighted average= 75.8

$ 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