DeepSeek V3.2 Speciale
Dec 2025
DeepSeek V3.2 SpecialeCompare inon
Benchmarks of DeepSeek V3.2 Speciale24
BenchmarkByScoreRank
Overall1
Knowledge4
Math9
Science2
Coding1
Writing2
Documents1
Legacy · saturated or no longer reported4
No benchmark matches
—per task?
self-host onlyprice
—output speed?
128Kcontext
Rankingin thinking mode?
Alternatives?
ModelScoreCostPer taskWhat changes
BetterGPT-6 Lunalow+49%—$0.005Function calling only with none · Chat Completions+3
OpenMiMo v2.6 Prothinking×3.2—$0.13Memory (4-bit): 411 GB → 615 GB+4BetterGPT-5.6 Lunalow+45%—$0.01BetterMiMo v2.5thinkingOpen+51%—$0.019
BetterClaude Haiku 5.5low×2—$0.024Remove sampling parameters+5ModelScoreCostPer taskWhat changes
Nothing here yet.
ModelScoreCostPer taskWhat changes
GPT-6 Lunalow+49%—$0.005Function calling only with none · Chat Completions+3GPT-5.6 Lunalow+45%—$0.01MiMo v2.5thinkingOpen+51%—$0.019Claude Haiku 5.5low×2—$0.024Remove sampling parameters+5GPT-5 Minihigh+16%—$0.053ModelScoreCostPer taskWhat changes
MiMo v2.6 ProthinkingOpen×3.2—$0.13Memory (4-bit): 411 GB → 615 GB+4DeepSeek V4.1 Flash 1017maxOpen×3.1—$0.26Memory (4-bit): 411 GB → 458 GB+5
GLM 5.3maxOpen×3.1—$2.01Memory (4-bit): 411 GB → 452 GB+4
Step 5 PreviewthinkingOpen×3—$1.03web_search not available · Chat Completions+2GLM 5.3 FlashmaxOpen×2.9—$0.28Defaults to max · Self-hosted+4ModelScoreCostPer taskWhat changes
DeepSeek V4.1 Flash 1017maxOpen×3.1—$0.26Memory (4-bit): 411 GB → 458 GB+5DeepSeek V4 Pro 0813non-thinkingOpen+41%—$0.36Memory (4-bit): 411 GB → 990 GB+5DeepSeek V3.1 Terminusnon-thinkingOpensame—$0.27 / $0.95DeepSeek V3.1non-thinkingOpensame—$0.25 / $0.95R1 0528thinkingOpen-9%—$0.39 / $1.63DeepSeek V3 0324non-thinkingOpen-33%—$0.021DeepSeek V3non-thinkingOpen-41%—$0.02DeepSeek R1 0528 Qwen3 8BthinkingOpen-44%—$0.06 / $0.09Context: 128K → 33KAll DeepSeek models →Run it yourself
estimateFits on✗ 8 GB (phone, laptop)✗ 16 GB✗ 24 GB GPU✗ 48 GB GPU✗ 80 GB GPU
LicenceMITfree for commercial use
Benchmarks
all categoriesChat & Knowledge
Humanity's Last Exam28.7#86
AA Omniscience Index-17.7#115
CritPt (physics)7.4#71
AA Intelligence Index 4.3.2≈14.5#150
AA Omniscience · Accuracy37.8#53
AIME 2025lab96
HMMT February 2025lab99.2
HMMT November 2025lab94.4
IMOAnswerBenchlab84.5
HMMT 2025 chart edition unstatedlab99.2
GPQA Diamond87.1#58
AIME 2025 · AA archived96.7#3
Coding
LiveCodeBenchlab88.7
Codeforces Ratinglab2,701
Agents & Tool Use
HLE without toolslab30.6
τ²-bench · AA0#232
Writing & Language
IFBench (instructions)63.9#74
FAQ
How much does DeepSeek V3.2 Speciale cost?›
There is no per-unit API price: self-host only.
Is DeepSeek V3.2 Speciale free to use?›
There is no free API. The weights are open, so you can run it yourself.
How good is DeepSeek V3.2 Speciale, and what is it best at?›
#86 of 334 in Chat & Knowledge, #44 of 239 in Coding, #229 of 242 in Agents & Tool Use, #66 of 95 in Writing & Language. Strongest at: Science.
Is DeepSeek V3.2 Speciale better than Gemini 2.5 Pro Preview (May' 25)?›
On AA Intelligence Index 4.3.2, DeepSeek V3.2 Speciale scores 14.5 vs 14.5 — about the same.
What is the context window of DeepSeek V3.2 Speciale?›
128K tokens.
When was DeepSeek V3.2 Speciale released? Is there a newer version?›
Released 2025-12-01. It is the current version in its line.
Is DeepSeek V3.2 Speciale open source?›
Yes, the weights are open (MIT licence).
Can I run DeepSeek V3.2 Speciale locally?›
The weights are open and can be self-hosted; hardware needs are not confirmed yet.
Can DeepSeek V3.2 Speciale read images?›
No, it accepts text only.
What is DeepSeek V3.2 Speciale?›
DeepSeek V3.2 Speciale is an open-weights language model from DeepSeek, released Dec 2025.
What are the best alternatives to DeepSeek V3.2 Speciale?›
GPT-6 Luna (better), MiMo v2.6 Pro (open weights).