Inference
EN
DeepSeek

DeepSeek V3.2

DeepSeek V3.2 is a standard model from DeepSeek. 23 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Standard
  • Input: text
  • Released 2025-12
RankedEnough results for an index.
IndexAll categories, out of 100.44.2±4.9Median 51.0
CoverageIndex weight that has results.85%
SpeedOutput tokens per second.19/sMedian 67/s
First tokenSeconds to the first output token.1.0 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.26Median $0.60
Output / 1MUS dollars per 1M output tokens.$0.38Median $2.55
ContextMaximum tokens in one request.164KMedian 262K
EloLMArena rating and rank.1425±3 #109Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
deepseek/deepseek-v3.2
Max outputMaximum output tokens per request.
147K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.13
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
6 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Deep InfraHost$0.26$0.38164K
OpenRouterRouter$0.26$0.38164K
SiliconFlowHost$0.27$0.42164K
ChutesHost$1.00$1.00131K
DigitalOceanCloud$0.50$1.60164K
Amazon BedrockCloud$0.62$1.85164K

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

685.36B parameters
16-bit
1440.3 GB
8-bit
765.6 GB
4-bit
437.3 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.6 counted43.1167 / 218
CodingCode writing and repair.4 counted48.8146 / 209
ReasoningLogic problems and puzzles.4 counted42.9150 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.4 counted43.2179 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.1 counted34.8—
MathMath problems.4 counted46.896 / 168

Results

23 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Terminal-Bench 1.0Agentic50.0% ±5.610 / 47
τ²-bench TelecomAgenticAbove fitFar above this model's index.96.2%4 / 18
SWE-bench Verified · SWE-bench/experimentsCoding70.0%16 / 65
τ²-bench RetailAgentic81.1%9 / 23
FrontierMath-2025-02-28-PrivateMath22.1% ±4.127 / 69
GPQA diamond · Epoch AIKnowledge77.3% ±2.096 / 217
τ²-Bench Tool-Agent-User EvaluationAgentic78.9%71 / 149
τ²-bench AirlineAgentic63.8%11 / 22
IOI v1Coding14.4% ±4.531 / 61
OTIS Mock AIME 2024-2025Math68.4% ±7.5103 / 200
Artificial Analysis IFBenchInstruction49.0%80 / 146
MMLU ProKnowledge83.1% ±0.474 / 135
ARC-AGI-1 (semi-private)Reasoning57.0%71 / 126
GPQA Diamond · Vals AIKnowledge76.3% ±2.278 / 135
AIMEMath64.8% ±2.055 / 95
Critical Physics TasksReasoning0.9%124 / 214
Artificial Analysis GPQA DiamondKnowledge75.1%118 / 199
Artificial Analysis Humanity's Last ExamKnowledge11.2%138 / 222
Terminal-Bench 2.0 · Vals AIAgentic34.8% ±5.142 / 66
FrontierMath-Tier-4-2025-07-01-PrivateMath2.1% ±2.135 / 55
LiveCodeBenchCoding69.9% ±1.291 / 142
ARC-AGI-2 (semi-private)Reasoning4.0%86 / 131
Artificial Analysis Omniscience AccuracyKnowledge24.0%110 / 164
Artificial Analysis Long Context ReasoningReasoning45.7%163 / 217
VITA-BenchAgentic18.5%11 / 13
React Native EvalsCoding71.5%14 / 16
Gert Labs Composite Game BenchmarkAgentic29.6%49 / 52
Claw-EvalAgentic40.2%41 / 42
MGSMMultilingual90.9% ±1.831 / 74
MedQAHealthcare89.5% ±0.351 / 95
ARC-AGI-1 (public eval)Reasoning61.6%67 / 116
ARC-AGI-2 (public eval)Reasoning3.9%76 / 122
CorpFin v2Finance47.9% ±1.0110 / 131
LegalBenchLegal66.0% ±0.4135 / 149
SWE-bench Multilingual · swebench.comMultilingual59.0%11 / 12
SWE-bench Multilingual · SWE-bench/experimentsCoding59.0%13 / 14
Finance Agent v1.1Finance0.0%51 / 51
SWE-RebenchCoding60.9%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseSierra ResearchMITVals AINo licence statedEpoch AICC BYSWE-bench teamNo licence statedARC Prize FoundationTerms restrict reusemodels.devMIT

History

5 releases
Index45.1 → 44.2
23 Oct 202612 Nov 2026
Output / 1M$0.42 → $0.38
23 Oct 202612 Nov 2026

Questions

How much does DeepSeek V3.2 cost?
$0.26 per 1M input tokens, $0.38 per 1M output tokens.
How good is DeepSeek V3.2?
Its index is 44.2 out of 100. That is place 185 of 283 ranked models.
What is DeepSeek V3.2 best at?
Math: 46.8 out of 100, place 96 of 168.
How fast is DeepSeek V3.2?
19 output tokens per second, 1.0 s to the first token.
What is the context window of DeepSeek V3.2?
164K tokens.
Is DeepSeek V3.2 open source?
Yes. Its weights are published.
When was DeepSeek V3.2 released?
December 2025.

Badge

DeepSeek V3.2 on Inference360: place 185 of 283, index 44.2
HTML
<a href="https://inference360.ai/models/deepseek-v3-2"><img src="https://inference360.ai/badges/deepseek-v3-2.svg" alt="DeepSeek V3.2 on Inference360: place 185 of 283, index 44.2" height="20"></a>
Markdown
[![DeepSeek V3.2 on Inference360: place 185 of 283, index 44.2](https://inference360.ai/badges/deepseek-v3-2.svg)](https://inference360.ai/models/deepseek-v3-2)

Compare with

Qwen3.5-4B44.3Grok 3 Mini44.3Claude Opus 443.9Claude Haiku 4.5 Thinking43.5DeepSeek V3.2 (Thinking)49.4

Lower price

dots3-note Preview63.4 · FreeApodex 1.1 Mini55.7 · FreeNemotron 3 Ultra50.8 · Free

More from DeepSeek

All models
DeepSeek V3.2 (Thinking)49.4DeepSeekMath V2UnrankedDeepSeek V3.2 ExpUnrankedDeepSeek V3.1 TerminusUnrankedDeepSeek V3.1 (Reasoning)43.0